10. AI for Science (AI4Science)

When deep learning and large models move from "recommendation ads and natural language understanding" toward scientific problems themselves, the goal is no longer just predicting a metric or performing classification, but genuinely participating in discovering patterns, designing experiments, and accelerating simulation and reasoning. AI4Science seeks to combine "statistical pattern recognition" with "physical laws / biochemical principles / mathematical structures," enabling models to serve as "programmable scientific assistants" in molecular design, protein engineering, materials discovery, physics simulation, mathematical reasoning, and beyond.

In engineering practice, this layer connects on one end to "traditional scientific infrastructure" such as quantum chemistry software, molecular dynamics (MD), CFD/FEA simulators, automated theorem provers, literature databases, and robotic labs, and on the other end to the real-world R&D workflows of pharmaceutical companies, materials enterprises, energy firms, and research institutions. The following discussion unfolds from three perspectives — scenarios, principles, and models — and further subdivides several key directions.

• Scenarios
  • Molecular and drug design: Starting from massive libraries of small molecules or fragments, predict properties and ADMET, design candidate drugs targeting specific targets, and narrow the experimental space through virtual screening and multi-objective optimization.
  • Protein and biological structure modeling: Predict the 3D structures of proteins and complexes, assist in antibody, enzyme, and protein drug design, and evaluate the impact of mutations on function and stability.
  • Physics simulation and engineering design: Use deep surrogate models to accelerate expensive simulations such as CFD / FEA / molecular dynamics, providing rapid evaluation and optimization tools for aerospace, automotive, energy, and other fields.
  • Materials discovery and crystal design: Conduct virtual screening and inverse design across vast chemical / materials spaces, accelerating the R&D of key materials such as batteries, photovoltaics, catalysts, and alloys.
  • Mathematics and symbolic reasoning: Perform automated theorem proving, symbolic computation, and equation solving within formal systems, enhancing the rigorous reasoning capabilities of large models in mathematical problems and engineering derivations.
  • Scientific workflows and automated experimentation: Connect with literature, databases, and automated experimental platforms to build "Self‑Driving Labs," enabling models to participate in experimental design, execution, and result analysis.
• Principles
  • Structured representation and graph modeling: Use graphs, crystal graphs, molecular graphs, and other structures to represent complex objects, modeling geometric and topological relationships on graph neural networks or E(3)-equivariant networks.
  • Physics / chemistry inductive biases: Incorporate physical priors into model architectures and loss functions through conservation laws, symmetries (translation / rotation / reflection), PDE constraints (PINN), energy potential functions, and other means.
  • Generation and inverse design: Use generative modeling methods such as VAE, GAN, Diffusion, and RL to support inferring structures from "target properties / constraints," enabling inverse design of molecules, materials, and structures.
  • Surrogate models and multi-scale coupling: Use deep surrogate models to approximate expensive quantum chemistry / continuum / structural mechanics simulations, and stitch together micro–meso–macro models to achieve multi-scale modeling.
  • Tool augmentation and agent workflows: Combine LLMs with simulators, symbolic calculators, automated theorem provers, literature retrieval systems, and experimental robots to build agents capable of autonomously planning and executing scientific tasks.
• Models
  • Molecular and materials representation models: E(3)-equivariant networks and graph networks such as SchNet, DimeNet, PhysNet, CGCNN, MEGNet, ALIGNN; molecular language models such as ChemBERTa, MolBERT, and MoleculeSTM.
  • Structural biology models: AlphaFold / AlphaFold2 / AlphaFold3, RoseTTAFold, OpenFold, ProteinMPNN, ESM‑IF, the ESM series of protein language models and structure generation models.
  • Physics simulation and operator learning: PINN, DeepONet, Fourier Neural Operator (FNO) and the Neural Operator family, DeepMD, NequIP, and other potential energy surface and operator learning models.
  • Mathematics and symbolic reasoning models: Math/proof-specialized models such as Minerva, Gödel, GPT‑f, and Lean‑Dojo, as well as tool-augmented systems combining LLMs with SymPy / Mathematica / Lean / Coq.
  • Scientific agents and workflow systems: Combining retrieval, code generation, simulation invocation, and experimental control interfaces to create "AI scientific assistants" and self-driving experimental platforms tailored for pharmaceuticals, materials, physics, chemistry, and other domains.

Starting from this layer, traditional scientific computing becomes deeply intertwined with deep learning and large models: both respecting the strict constraints of physics / chemistry / biology / mathematics, and leveraging the powerful data-driven fitting capabilities to improve efficiency. The ultimate goal is for AI to become a "collaborator" in scientific research, not merely a predictive black box.

---

10.1 Molecular Modeling & Drug Discovery

In traditional drug development, going from target discovery to clinical trials often takes 10+ years and billions of dollars, with a huge portion of that time and cost consumed in the early stages of molecular design, property prediction, and virtual screening. AI-driven molecular modeling and drug discovery aims to accelerate this process using data-driven + generative modeling: starting from structures or textual descriptions, predict molecular properties and ADMET, design candidate compounds targeting specific targets, and significantly reduce the burden of wet-lab experiments through multi-objective optimization and virtual screening.

This direction connects on one end to data sources such as quantum chemistry software (DFT, ab initio), bioactivity assays, and HTS (High‑Throughput Screening), and on the other end to internal Small Molecule Design platforms, property prediction SaaS, and materials / chemicals design tools within pharmaceutical companies. The following unfolds from three dimensions — scenarios, principles, and models.

• Scenarios
  • Early virtual screening and hit discovery: Facing virtual molecular libraries on the scale of millions to billions, rapidly predict activity / ADMET via AI, rank candidate molecules, and filter out a small number of high-value hits for experimental validation.
  • Molecular property and ADMET evaluation: During the lead optimization stage, continuously predict metrics such as solubility, toxicity, metabolic stability, and oral bioavailability, providing references for pharmacokinetic and safety assessment.
  • Target-oriented molecular generation: Given protein target information (pocket characteristics, known ligands) or target property constraints, automatically generate candidate small molecules with structural diversity, high activity, and synthetic feasibility.
  • Materials and chemicals molecular design: For non-pharmaceutical scenarios such as coatings, solvents, electrolytes, and surfactants, design formula molecules that meet specific physical properties (viscosity, polarity, interfacial energy, etc.).
• Principles
  • Molecular representation and property prediction:
    • Structural representation: Common forms include SMILES sequences, molecular graphs (atoms as nodes, bonds as edges), 3D coordinates, and quantum features; models need to extract generalizable semantic and geometric information from these representations.
    • Property prediction: Through GNNs (GCN, GAT, MPNN) or 3D-equivariant networks (SchNet, DimeNet, PhysNet, etc.), learn quantum properties such as energy, dipole moment, and orbital energy levels from molecular graphs or 3D structures, as well as ADMET attributes such as solubility, LogP, toxicity, and metabolic stability.
    • Representation learning and pretraining: Conduct masked prediction, contrastive learning, or autoregressive pretraining on large-scale molecular libraries (e.g., ZINC, ChEMBL, PubChem) to obtain transferable universal molecular representations, providing features for downstream QSAR / ADMET tasks.
  • Structure generation and molecular optimization:
    • Generative modeling: Use generative models such as VAE, GAN, Flow, and Diffusion to sample new molecules in SMILES or molecular graph space, ensuring chemical structure validity (valence, ring structures, etc.) and diversity.
    • Conditional generation: Introduce condition vectors (target activity, physicochemical properties, structural fragments, target pocket descriptions, etc.) to generate candidate molecules under given constraints, enabling property-oriented or fragment-completion-based design.
    • Multi-objective optimization and RL: Through reinforcement learning (e.g., MolDQN) perform "editing" operations (adding atoms, changing bonds, replacing fragments) in molecular space, balancing multiple objectives such as activity, toxicity, synthetic feasibility, and patent avoidance.
  • Protein–small molecule interaction modeling:
    • Binding site and scoring functions: Model the spatial relationship between protein pockets and ligands through 3D convolution / graph networks / interaction graphs, predicting binding sites and binding affinity.
    • Docking and binding pose prediction: Combine conformational search in docking with deep models, using deep scoring functions or diffusion-based generation to predict stable conformations, improving docking accuracy and reducing computational cost.
• Models
  • Molecular representation models:
    • GNNs and 3D networks: 3D-equivariant models such as DimeNet / DimeNet++, SchNet, and PhysNet that consider angles / distances; general graph neural networks such as GCN / GAT / MPNN, suitable for property prediction and QSAR.
    • SMILES-based Transformers: Treat molecules as "chemical language sentences," using Transformers for autoregressive or masked language modeling, providing sequence representations for generation and property prediction.
  • Generation and optimization models:
    • Graph generation models: GraphVAE, Junction Tree VAE, GraphAF, etc., generate molecules in graph / fragment space, emphasizing structural validity and interpretability (fragment-level construction).
    • Diffusion models: Diffusion for Molecules generates new molecules or conformations by adding/removing noise in graph or 3D structure space, and can be combined with condition vectors for customized generation.
    • Reinforcement learning optimization: RL-based methods such as MolDQN treat molecular optimization as a sequential decision-making problem in a "molecular editing" state space, encoding multi-objective metrics via reward functions.
  • Molecular large models and multimodal directions:
    • Molecular language models: ChemBERTa, MolBERT, etc., pretrained on large-scale SMILES corpora, supporting zero-shot or few-shot transfer to downstream tasks.
    • Multimodal molecular models: MoleculeSTM and others integrate structure (graph / 3D), textual descriptions (synthesis routes, literature abstracts), and molecular properties, enabling cross-modal retrieval and joint prediction.
  • Products and application forms:
    • Early-stage drug screening platforms and internal Small Molecule Design platforms for pharmaceutical companies, providing integrated capabilities including virtual screening, molecular generation, and ADMET prediction.
    • Property prediction SaaS for R&D personnel: quickly query molecular properties, ADMET, molecular similarity, etc., via web or API.
    • Molecular-level design tools for materials and chemicals design, used for custom development of molecular systems such as coatings, solvents, and electrolytes.

Starting from this sub-direction, the drug design workflow is moving from "experts + high-throughput experimentation" toward a closed loop of "experts + models + automated experimentation," where AI not only gives scores but gradually participates in the full cycle from "proposing ideas" to "generating candidates" to "screening and optimization."

10.1.1 Molecular Representation and Property / ADMET Prediction

In drug and materials R&D, a fundamental capability is: given a molecule, quickly and accurately predict its properties and behavior, including quantum chemical properties (energy, orbitals, dipole moment), physicochemical properties (solubility, LogP), and pharmacokinetic / toxicity-related ADMET metrics. The essence of this problem is how to learn, from different forms of molecular representation, a representation that both conforms to chemical principles and possesses generalization ability.

• At the molecular representation level, common representations include:
  • SMILES / SELFIES and other strings: Treat molecules as sequences, naturally suited for language modeling with RNNs / Transformers.
  • Molecular graph representation: Atoms as nodes, bonds as edges, with nodes and edges carrying features such as type, valence, and aromaticity; suitable for modeling neighborhoods and topology using GNNs, MPNNs, etc.
  • 3D geometric representation: 3D coordinates, bond angles, dihedral angles, and other information obtained from quantum chemistry or force field optimization, providing the basis for E(3)-equivariant networks to capture spatial structure.
• At the property and ADMET prediction level, target tasks include:
  • Small molecule quantum property prediction: energy, dipole moment, HOMO/LUMO energy levels, etc., used to replace expensive DFT / ab initio calculations.
  • QSAR / activity prediction: predicting a compound's activity (IC50, Ki), selectivity, etc., against specific targets, for screening potential candidates.
  • ADMET-related metrics: solubility, permeability, toxicity, metabolic stability, CYP inhibition, etc., which are key to assessing drug developability.

Typical model pathways: use DimeNet / SchNet / PhysNet / GNNs to extract high-dimensional representations from molecular structures, then simultaneously predict multiple properties via multi-task learning; conduct pretraining on large-scale public or internal corporate data to improve modeling capability in small-data scenarios. Externally, these are offered as ADMET prediction SaaS or internal platform APIs, providing project teams with rapid "virtual experiment" capabilities.

10.1.2 Structure Generation and Molecular Optimization: From SMILES / Graph to Drug Candidates

With reliable molecular representation and property prediction models in place, the next goal is to actively generate "better" molecules: no longer just evaluating given compounds, but directly designing new candidate molecules around targets and property constraints. This direction is commonly referred to as molecular generation and molecular optimization.

In terms of structure generation, research and engineering practice center on three main pathways:

1. SMILES-based sequence generation Treat molecules as strings, using VAE, GAN, or autoregressive Transformers to sample new structures in SMILES space; ensure chemical validity through grammatical constraints (e.g., SELFIES) or post-processing.
2. Graph / fragment-based generation Models such as GraphVAE, Junction Tree VAE, and GraphAF construct structures directly at the molecular graph or primitive fragment (fragment / motif) level, which is closer to chemical synthesis thinking and facilitates control over rings, groups, and scaffold structures.
3. Diffusion and 3D generation Methods such as Diffusion for Molecules perform diffusion and denoising in graph or 3D coordinate space, simultaneously considering spatial conformation, suitable for generating ligands or material units that are sensitive to 3D shape.

In terms of molecular optimization, the key is introducing objectives and constraints:

• Conditional generation: Input target activity, physicochemical properties, or fragment anchors as condition vectors into the model, biasing generation toward satisfying these conditions.
• Reinforcement learning and multi-objective optimization: Use property prediction models as the "environment," employ RL for sequential decision-making in molecular space (e.g., MolDQN), setting rewards and penalties on multi-dimensional metrics such as activity, toxicity, synthetic feasibility, and patent risk to achieve multi-objective trade-offs.
• Synthetic feasibility and chemical priors: Incorporate synthesis route prediction models and synthetic complexity metrics (e.g., SA score) during generation and optimization to avoid producing structures that are difficult to synthesize or unstable.

In terms of productization, these models are often packaged into internal "AI drug design platforms" within pharmaceutical companies: given a target, known lead structures, and optimization directions, the platform automatically proposes batches of candidate molecules; project teams then progressively screen and iterate in combination with experimental, patent, and commercial considerations, forming a "model–experiment–model" closed-loop optimization.

10.2 Protein & Structural Biology

In life sciences, structure determines function is a near-dogmatic principle: how a protein folds into a three-dimensional structure and how it assembles into complexes with other molecules directly determines its functional behavior in cells. Traditional structural determination relies on experimental techniques such as X‑ray crystallography, NMR, and cryo‑EM, which are time-consuming, expensive, and have massive blind spots where "difficult to crystallize, difficult to resolve" applies. Deep learning models represented by AlphaFold have dramatically advanced the ability to go "directly from sequence to structure," making it possible to obtain high-quality structures at the genome-wide scale.

This direction connects on one end to sequence and structure databases such as UniProt / PDB, omics experiments, and structural genomics projects, and on the other end to structure design and analysis platforms in industries such as biopharmaceuticals, synthetic biology, and enzyme engineering. The following unfolds from three perspectives — scenarios, principles, and models — and further breaks down key sub-directions.

• Scenarios
  • Target structure annotation and screening: Predict structures for large numbers of proteins at the genome level, aiding target discovery, functional annotation, and pathway analysis; combine with variant information to assess potential pathogenic mechanisms.
  • Antibody / protein drug design: Perform fine-grained modeling and design of key regions such as antibody variable regions (CDRs) and receptor binding domains, optimizing affinity, specificity, and immunogenicity.
  • Enzyme and biocatalysis design: Based on enzyme 3D structures and active site environments, design mutation and variant libraries to improve catalytic efficiency, substrate scope, and stability.
  • Complex and interaction studies: Predict protein–protein, protein–nucleic acid, and protein–small molecule complex structures, resolve interface interaction patterns, and provide a foundation for drug design and signaling pathway modeling.
  • Mutation effect and drug resistance analysis: Evaluate the impact of natural variants or artificial mutations on structural stability, function, and ligand binding; analyze the structural basis of drug resistance mutations.
• Principles
  • Protein structure prediction:
    • Sequence → structure: Starting from amino acid sequences (single sequence or including multiple sequence alignment MSA), model pairwise geometric constraints between residues (distances, angles, contact maps), then generate full-atom 3D structures via geometry reconstruction modules.
    • Coevolutionary signals: Use co‑mutation patterns (co‑evolution) among homologous sequences to infer potential residue contact relationships, providing strong priors for folding constraints.
    • Structure refinement and uncertainty estimation: Perform local refinement (relax, repack) on predicted structures, and output confidence scores (e.g., pLDDT, PAE) to guide the selection of "trusted regions" in downstream applications.
  • Complex and molecular assembly modeling:
    • Multi-chain joint modeling: Take multiple protein chains or protein + nucleic acid sequences as input, introduce chain identification and interface constraints, and directly output complete complex structures.
    • Interface prediction and assembly: Based on known monomer structures, predict the most likely interface configurations and assembly modes through graph models or diffusion models.
  • Protein design and mutation effect prediction:
    • Inverse folding: Given a target 3D backbone structure or topological constraints, generate amino acid sequences that can stably fold into that structure, enabling de novo protein design.
    • Mutation effect modeling: Combine protein language models with structural models to predict the impact of specific mutations on stability (ΔΔG), activity, or binding affinity, aiding directed evolution and variant screening.
• Models
  • Structure prediction:
    • AlphaFold / AlphaFold2 / AlphaFold3: Centered on attention mechanisms and geometric modules, predict high-precision protein structures from MSA, template structures, and sequence features, and output uncertainty estimates.
    • RoseTTAFold, OpenFold: Employ multi-track (sequence / pair / structure) representations and multi-scale attention mechanisms, providing foundational implementations for open-source and industrial deployment.
  • Complex and interface modeling:
    • AlphaFold‑Multimer: Directly model protein–protein complex structures in multi-chain scenarios, accounting for both monomer folding and interface interactions.
    • RFdiffusion: Based on diffusion models, generate or optimize protein backbones and complex interfaces in 3D space, enabling complex assembly and symmetric body design.
    • Methods such as DiffDock: In protein–small molecule systems, use diffusion or deep scoring functions to predict binding poses and binding modes.
  • Design and mutation models:
    • ProteinMPNN: Generate compatible sequences given a structure, used for stabilizing backbones and interface design.
    • ESM‑IF, ESMFold / ESM‑2 series: Language models pretrained on large-scale protein sequences, with the ability to infer structure, function, and mutation effects from sequences.
  • Products and applications:
    • Public cloud protein structure prediction services and databases (e.g., AlphaFold DB), providing large-scale structural annotations and download interfaces for research.
    • Internal structure design platforms at biopharmaceutical companies: integrating modules for protein structure prediction, antibody design, enzyme engineering, and protein–ligand docking.
    • Biotechnology SaaS: Providing tools for binding site prediction, interface thermodynamics evaluation, affinity and immunogenicity assessment, serving antibody drug and biologic development.

From this sub-direction onward, AI is not only "interpreting" naturally occurring protein structures but also "creating" entirely new protein and complex architectures, moving structural biology from the "passive measurement era" into the "active design era."

10.2.1 Protein Structure Prediction and Complex Assembly

Protein structure prediction is one of the most representative breakthroughs at the intersection of structural biology and AI. The core question is: can we predict 3D structures at near-experimental resolution from sequence alone, with little to no reliance on experimental data? In real-world applications, monomer structures are often just the starting point; more critically, how proteins assemble into complexes with other molecules.

In monomer structure prediction, the typical workflow includes:

1. Sequence / MSA encoding: Extract sequence features and mine coevolutionary signals through multiple sequence alignment.
2. Geometric constraint inference: Predict distance distributions, contact probabilities, and relative orientations between residue pairs, forming a "pseudo-measurement" geometric field.
3. Structure construction and iterative refinement: Build 3D structures under geometric constraints using structural modules (e.g., rotation-translation invariant blocks, internal coordinate updates), and iteratively refine multiple times to reduce geometric violations.
4. Uncertainty and quality assessment: Output per-residue confidence (pLDDT), residue-pair error estimates (PAE), and other metrics to inform subsequent modeling and screening.

In complex and assembly prediction, the problem further extends to "how multiple chains organize and interact in space":

• For protein–protein complexes, specialized multi-chain modeling strategies (e.g., AlphaFold‑Multimer) are typically used with multi-chain inputs to directly output assembly structures.
• For protein–nucleic acid / protein–small molecule systems, one approach is to first predict individual structures, then predict assembly modes through docking and interface scoring functions; another is to use diffusion models or joint modeling to directly generate complex conformations in 3D space.
• In multi-subunit, large assembly scenarios, it is also necessary to incorporate symmetry constraints, low-resolution EM density maps, and other information for hierarchical and multi-scale assembly.

In product practice, structure prediction and assembly are often packaged as cloud services or local toolchains, providing foundational structural information for protein function annotation, interaction network modeling, and drug target validation.

10.2.2 Protein Design and Mutation Effect Prediction: From Structure to Functional Regulation

After mastering the "sequence → structure" mapping, the next step is the inverse problem: given a structure or functional requirement, how do we design suitable protein sequences and mutation strategies? This is the core of protein design and mutation effect prediction.

In protein design, key tasks include:

• Inverse folding: Given a target backbone or overall topological structure, generate amino acid sequences that can stably fold into that structure. This process can be achieved through structure-conditioned generative models such as ProteinMPNN and ESM‑IF.
• Function-oriented design: While maintaining overall structural stability, perform targeted design on active sites, binding pockets, and interface regions to optimize affinity, specificity, and catalytic efficiency.
• Manufacturability and immunogenicity constraints: During sequence design, introduce constraints such as expression feasibility, post-translational modifications, and immunogenicity risk to ensure the practical viability of candidate sequences in biologic development.

In mutation effect prediction, the focus is on:

• Stability change (ΔΔG): Given a wild-type structure and mutation sites, predict the impact of single-point or multi-point mutations on folding stability, used for directed evolution and drug resistance mutation analysis.
• Activity and affinity changes: Combine structural and protein language models to evaluate the impact of mutations on enzymatic activity, ligand affinity, and signaling pathway regulation.
• Large-scale variant library design: Before in vivo / in vitro screening experiments, use models to pre-screen vast mutation spaces, retaining high-potential variants and reducing experimental costs.

At the engineering and product level, protein design and mutation effect prediction are often integrated as "structural design and optimization modules" within biopharmaceutical / synthetic biology companies: starting from candidate backbone structures, automatically propose multiple rounds of mutation and variant library design plans, forming a data-driven closed loop with high-throughput screening experiments.

10.3 Physics Simulation & Surrogate Modeling

In aerospace, automotive, civil engineering, energy, chemical, and other fields, high-fidelity simulation is a core part of design and verification. However, CFD (Computational Fluid Dynamics), FEA (Finite Element Analysis), molecular dynamics (MD), and various PDE solvers are often computationally expensive, making it difficult to support large-scale parameter sweeps, real-time control, or online optimization. AI-driven physics simulation and surrogate modeling seeks to use deep networks to approximate numerical solvers or operators themselves, achieving orders-of-magnitude acceleration while maintaining physical consistency and interpretability.

This direction connects on one end to traditional simulation software (ANSYS, Fluent, COMSOL, custom solvers), experimental measurements, and sensor data, and on the other end to engineering design platforms, autonomous driving and aerospace aerodynamic design, and chemical process simulation and optimization systems. The following unfolds from three perspectives — scenarios, principles, and models.

• Scenarios
  • Engineering simulation acceleration: Given geometry and operating conditions, use deep surrogate models to rapidly predict pressure fields, velocity fields, temperature fields, stress/strain distributions, etc., supporting multi-round design iteration and optimization.
  • Complex process simulation and process optimization: In process industries such as chemicals and energy, use ML to approximate mechanistic models or black-box process models, enabling rapid evaluation and real-time control.
  • Molecular / materials scale simulation: Use ML potential energy surfaces (Neural Network Potential) to replace expensive ab initio potential energy and force calculations, accelerating molecular dynamics and materials phase behavior simulation.
  • Multi-scale and cross-disciplinary coupling: Use deep surrogate models to stitch together micro–meso–macro models, building end-to-end multi-scale simulation and optimization chains.
• Principles
  • Surrogate models:
    • Learn the "input parameters → output fields / metrics" mapping from numerical simulation or experimental data, serving as an approximation of high-fidelity solvers.
    • In high-dimensional parameter spaces, combine active learning and Bayesian optimization to automatically select the most informative sample points for high-fidelity simulation or experimentation, continuously improving surrogate model quality.
  • Physics-Informed Neural Networks (PINN):
    • Write PDEs, initial / boundary conditions, and physical conservation laws into the loss function, using automatic differentiation to solve physical fields over continuous space.
    • Support forward problems (solving state fields) and inverse problems (inferring source terms, material parameters, etc., from sparse observations), particularly suited for complex geometries and boundaries that are difficult for traditional numerical methods.
  • Operator learning and Neural Operators:
    • Not just fitting "solutions under specific conditions," but learning mappings from functions to functions (operators), such as "boundary conditions / source terms → entire solution field."
    • Representative methods include Fourier Neural Operator (FNO), DeepONet, etc., which improve generalization across different mesh densities and geometries through frequency-domain transformations or specialized network architectures.
  • Multi-scale modeling:
    • Train effective parameters or constitutive relations at meso / macro levels on micro-scale simulation data, with deep surrogate models serving as the "scale bridging layer."
    • For complex materials, fluid-structure coupling, and multiphase flow problems, use deep models to transfer information between different scales and physics modules.
• Models
  • General physics neural networks:
    • PINN series: Solve by minimizing PDE residuals at spatiotemporal sampling points, applicable to Navier‑Stokes, Maxwell, elasticity, and other equations.
    • DeepONet, FNO, Neural Operator family: Directly learn "operator-level" approximations of PDE solvers, enabling fast inference across multiple operating conditions and geometries.
  • Molecular / materials scale potential energy models:
    • DeepMD, SchNet, NequIP, SpookyNet, etc.: Build high-precision ML potential energy surfaces, significantly accelerating force and energy calculations while maintaining near-ab initio accuracy.
    • Couple with traditional MD engines to achieve high-precision molecular dynamics for large systems and long timescales.
  • CFD / structural mechanics surrogate models:
    • U‑Net / UNet++ and other Encoder‑Decoder networks: Predict flow fields or temperature fields from geometry / boundary conditions on regular grids.
    • Graph neural networks on mesh: Perform message passing and updates on nodes / elements over unstructured meshes, suitable for complex geometries and multi-physics coupling scenarios.
    • Neural Operator for CFD: Generalize flow field predictions across different Reynolds numbers, inflow conditions, and geometric parameters.
  • Products and applications:
    • AI acceleration modules in industrial simulation software: Provide rapid estimation and sensitivity analysis capabilities on top of traditional solvers.
    • Chemical / energy process simulation and optimization platforms: Combine mechanistic models + surrogate models + optimization algorithms into integrated process optimization tools.
    • Autonomous driving / aerospace aerodynamic design: Conduct large-scale design variable sweeps and automatic shape optimization in aerodynamic shape design.

10.3.1 Surrogate Models and Physics-Informed Neural Networks (PINN)

Surrogate models and Physics-Informed Neural Networks (PINN) are two complementary paths for AI-enabled physics simulation: the former approximates simulation mappings from data, while the latter constructs learning objectives from physics.

In surrogate model scenarios, the typical workflow is:

1. Collect a batch of sample data (input parameters, boundary conditions, geometry → output physical quantities) through high-fidelity numerical simulation or experimentation.
2. Train a deep network (e.g., MLP, convolutional networks, GNNs, Neural Operators) to approximate this mapping function.
3. In design optimization, parameter sweeps, or real-time control, use the surrogate model in place of expensive solvers for rapid evaluation.

In PINN scenarios, the model no longer relies primarily on large amounts of supervised labels, but instead constructs the loss function by minimizing PDE residuals and boundary condition violations:

• At spatial / temporal sampling points, use the neural network to output physical quantities (e.g., velocity, pressure, displacement fields), and obtain gradients and derivatives via automatic differentiation.
• Substitute these derivatives into the PDE to form residuals, which together with the errors of boundary conditions and initial conditions constitute the total loss.
• Optimize to make PDE residuals and boundary errors as close to 0 as possible, thereby obtaining an approximate solution that satisfies the physical equations.

The two can be used in combination: when partial high-fidelity data is available, jointly constrain training with data error + physical residuals to improve accuracy and generalization. In engineering applications, PINN is particularly well-suited for inverse problems and data-driven modeling, such as inferring material parameters, source terms, or defect locations from sensor observations.

10.3.2 Neural Operators and Multi-Scale Physics Modeling

Neural Operators elevate physics modeling from "point-to-point / parameter-to-solution" mappings to the "function-to-function" level: they learn a unified operator approximation for "given a class of PDEs and boundary conditions, solve for the solution field," rather than a specific solution under a single operating condition. This opens new possibilities for generalization across multiple conditions, geometries, and mesh resolutions.

In operator learning, typical approaches are:

• Take functions (e.g., source terms, boundary conditions, material parameter fields) as input, and use networks (e.g., FNO, DeepONet) to output the entire solution field function.
• Train on samples across different meshes, parameters, and geometries, so the model learns the "common patterns" of PDE solvers.
• At deployment, simply provide new input functions (e.g., new boundary conditions, geometry) to rapidly infer approximate solution fields.

In multi-scale modeling scenarios:

• Train Neural Operators on large amounts of data generated at the micro scale (e.g., molecular dynamics, crystal plasticity) to learn the mapping between microstructure and macroscopic response.
• In macroscopic continuum models, use this mapping as a constitutive relation or effective parameter computation module to achieve micro–macro coupling.
• For complex systems such as fluid-structure coupling, multiphase flow, and reacting flow, separately model different physical fields and couple them through shared interface variables (e.g., fluxes, interface forces).

In engineering practice, Neural Operators are gradually moving from research prototypes to applications, becoming an important technical direction for "accelerated solvers + multi-scale bridging" in CFD, geophysics, climate modeling, and other scenarios.

10.4 Materials Science & Crystal Design

In materials science, a core contradiction is: the design space is nearly infinite, while the cost of experimentation and high-precision computation is extremely high. How to efficiently find candidate materials that meet specific performance requirements within the vast chemical and structural combination space is a key problem in new energy, electronics, structural materials, functional materials, and other fields. AI-driven materials discovery and crystal design, through graph neural networks, generative models, and high-throughput virtual screening, is progressively shifting "trial-and-error" R&D toward "data-driven + inverse design."

This direction connects on one end to materials databases such as Materials Project, OQMD, AFLOW and DFT / MD computation results, and on the other end to materials R&D platforms in application scenarios such as batteries, photovoltaics, catalysis, semiconductors, and alloys. The following unfolds from three perspectives — scenarios, principles, and models.

• Scenarios
  • Performance-oriented materials screening: Given crystal structures or chemical formulas, predict band structures, band gaps, carrier mobility, thermal / electrical / magnetic properties, etc., providing a basis for materials screening and combinatorial optimization.
  • New energy materials R&D: For systems such as battery electrolytes, electrode materials, solid-state ion conductors, photovoltaic absorber layers, and catalysts, predict ionic conductivity, stability, electrochemical windows, activity, etc.
  • High-throughput virtual screening (HTVS): In large-scale constructed candidate libraries, rapidly evaluate using ML models, filter out promising materials, then validate and calibrate with a small amount of DFT / experimentation.
  • Crystal structure and composition inverse design: Starting from target properties, reverse-search for crystal structure / composition combinations that satisfy performance and process constraints.
• Principles
  • Materials and crystal representation:
    • Represent periodic crystal structures as crystal graphs: nodes as atoms, edges as interatomic neighbor relationships, combined with lattice parameters and space group information.
    • For amorphous or complex multi-phase materials, use local environment descriptors (e.g., SOAP), Voronoi features, or multi-scale graph structures to represent their microstructure.
  • Property prediction:
    • Perform convolution / message passing on crystal graphs using GNN models such as CGCNN, MEGNet, and ALIGNN to predict energy, band gap, elastic modulus, thermal conductivity, etc.
    • Use embeddings based on literature and chemical formulas such as Mat2Vec to achieve transfer learning and zero-shot estimation in low-data scenarios.
  • High-throughput virtual screening:
    • Construct candidate libraries (via combinatorial enumeration, structure generation, empirical rules, etc.) → rapidly predict properties using ML models → select top candidates for DFT or experimental calibration → update models and screening strategies, forming an active learning closed loop.
  • Generation and inverse design:
    • Use diffusion models, VAEs, or GNN generative models to sample new structures in crystal structure space, with constraints such as composition, space group, and density.
    • Combine surrogate models with Bayesian optimization to search for suitable structure / composition combinations from target properties, achieving inverse design.
• Models
  • Representation and prediction:
    • CGCNN (Crystal Graph Convolutional Neural Network): Performs convolution on crystal graphs for predicting properties of inorganic materials such as energy and band gap.
    • MEGNet, ALIGNN: Integrate graph structure with edge / angle information, offering stronger generalization and accuracy across multiple material families.
    • Mat2Vec + lightweight ML: Rapidly train small models for specific property prediction by vectorizing chemical formulas and elemental information.
  • Generation and inverse design:
    • Diffusion for Crystals: Perform diffusion / denoising in the high-dimensional space composed of lattice parameters and atomic positions, generating crystal structures that satisfy certain constraints.
    • GNN‑based Generative Models: Search from random initialization toward structures near target properties by progressively adding/modifying atoms and bonds or manipulating the lattice.
    • Surrogate + Bayesian Optimization: Use ML models as approximate black boxes for "structure → property" mapping, performing Bayesian optimization on top to find optimal structures or compositions.
  • Data platforms and toolchains:
    • Materials Project, OQMD, AFLOW: Provide large amounts of structure and DFT computation data, forming the foundation for training and evaluating materials ML models.
    • Internal corporate materials databases and models: Combine company experimental data and process information to build domain-specific materials AI design platforms.
  • Products and applications:
    • New energy materials R&D acceleration platforms: Provide integrated property prediction, HTVS, and inverse design capabilities for battery, electrocatalysis, photovoltaic, and other teams.
    • Virtual screening software and SaaS: Provide digital screening tools for alloys, semiconductors, functional ceramics, etc., reducing early trial-and-error costs.
    • AI design tools within materials companies: Interface with Laboratory Information Management Systems (LIMS) and production line data, forming a closed loop from "model → experiment → production."

10.4.1 Materials Property Prediction and High-Throughput Virtual Screening (HTVS)

In the materials R&D workflow, fast and reliable property prediction is a foundational capability: given a candidate structure or composition, can we roughly determine whether it is worth further exploration without performing expensive DFT / experimentation? Property prediction models based on GNNs and materials databases make high-throughput virtual screening possible.

At the property prediction level:

• Use crystal graphs to represent periodic structures, learning interactions between atoms and their neighborhoods through models such as CGCNN, MEGNet, and ALIGNN.
• Conduct single-task or multi-task training for different tasks (energy, band gap, elastic constants, thermal conductivity, electrical conductivity, magnetism, etc.), achieving prediction performance close to DFT accuracy on datasets such as Materials Project.
• In industrial scenarios, often combine with internal experimental data for retraining or domain adaptation to improve suitability for specific material families and process conditions.

In high-throughput virtual screening (HTVS) scenarios, the typical workflow is:

1. Construct large-scale candidate libraries (combinatorial enumeration, structure generation, or expansion from existing databases).
2. Use ML models to rapidly predict the target properties and auxiliary properties (stability, safety, cost-related metrics, etc.) of each candidate.
3. Rank by target properties and multiple constraints, selecting Top‑K candidates for high-fidelity DFT computation or experimental validation.
4. Feed validation results back to the model, updating parameters and uncertainty estimates, forming an active learning closed loop of "screening–validation–re-screening."

This workflow has entered practical use in multiple domains including battery materials, photovoltaic absorber layers, catalysts, and structural materials, becoming a "front-end screening engine" for materials R&D teams.

10.4.2 Crystal Generation and Inverse Design: From Target Properties to Candidate Structures

With reliable property prediction and HTVS capabilities in place, the next goal is to directly propose new crystal structures and composition candidates from target properties and constraints — i.e., materials inverse design and generation.

In crystal generation, key questions include:

• How to generate physically reasonable lattices and atomic arrangements under periodic constraints?
• How to explicitly or implicitly impose constraints such as composition, symmetry, and density during generation?
• How to ensure generated structures remain stable after simple relaxation?

To address these, research and engineering practice commonly adopt:

• Diffusion for Crystals: Add/remove noise in the joint space of lattice parameters + atomic positions, achieving progressive generation from random initialization to structural samples, with target properties and composition constraints incorporated into the noise process or condition vectors.
• GNN‑based Generative Models: Gradually add atoms and connections on graph structures, or edit existing structures, to generate candidate structures that satisfy constraints.

In inverse design, it is typically combined with surrogate models and optimization methods:

• Treat property prediction models as black-box functions for "structure → property."
• Explore the structural space through Bayesian optimization, evolutionary algorithms, or RL, progressively bringing predicted properties closer to target values while satisfying constraints such as stability, safety, and cost.
• Validate the searched candidate structures through DFT / experimentation, and use the results to update surrogate models and search strategies.

In engineering applications, inverse design modules are often integrated into materials AI platforms, providing R&D personnel with an interactive interface for "set target properties → system automatically proposes candidate structures," significantly improving the efficiency of new materials exploration.

10.5 Mathematics & Symbolic Reasoning

Mathematics is a highly formalized, precisely verifiable language, which gives it both "extremely high difficulty" and "potentially enormous returns" in the AI era. On one hand, complex theorem proving and higher-order reasoning place extremely high demands on model capabilities; on the other hand, the results of mathematical reasoning and symbolic computation can be rigorously verified, making them naturally suited for collaboration with programmatic tools. The goal of AI in mathematics and symbolic reasoning is to build models capable of reliable reasoning and computation within formal systems, and to integrate them into education, research, and engineering applications.

This direction connects on one end to interactive theorem provers such as Lean / Coq / Isabelle, computer algebra systems (CAS) such as SymPy / Mathematica / Maple, and large mathematical problem banks and literature corpora; on the other end to mathematics education products, research assistance tools, and formula derivation and risk analysis needs in engineering / finance and other fields. The following unfolds from three perspectives — scenarios, principles, and models.

• Scenarios
  • Automated theorem proving and assisted proving: Automatically generate theorem proofs within formal systems, or produce readable proof drafts for human review and refinement.
  • Expression manipulation and symbolic computation: Automatically simplify expressions, differentiate, integrate, perform series expansions, transformations, and equation solving, providing symbolic tools for engineering modeling and financial risk analysis.
  • Math problem understanding and solution step generation: Extract structured representations from problems in natural language or images, produce rigorous, verifiable solution steps, serving education and training scenarios.
  • Mathematical reasoning capability enhancement: Through math-specific fine-tuning and tool augmentation, improve large models' multi-step reasoning and rigor in arithmetic, algebra, geometry, combinatorics, and other domains.
• Principles
  • Formal systems and search:
    • Within systems such as Lean / Coq / Isabelle, mathematical objects and theorems are formalized as terms and types, and the proof process corresponds to constructing proof trees under rule constraints.
    • Proof search can be viewed as "finding a path satisfying constraints in an extremely large state space," suitable for methods such as reinforcement learning, MCTS (Monte Carlo Tree Search), and policy / value networks.
  • Neural–symbolic collaboration:
    • LLMs are responsible for extracting problem structures and solution approaches from natural language or unstructured inputs, translating them into symbolic expressions (e.g., SymPy code, Lean proof scripts).
    • Computer algebra systems and theorem provers are responsible for executing rigorous symbolic computation and formal verification, validating and correcting LLM outputs.
  • Mathematical reasoning capability improvement:
    • Through targeted pretraining or fine-tuning on large-scale mathematical texts and problem banks (e.g., Minerva, Gödel), enhance models' understanding of mathematical language and mastery of reasoning styles.
    • Adopt Tool‑Augmented LLM frameworks, incorporating symbolic solvers, numerical computation libraries, plotting tools, and theorem provers as external tools, enabling models to learn to "call tools" rather than "memorize results" in complex reasoning.
• Models
  • Automated theorem proving:
    • AlphaZero‑style provers: View the proof process as a game, using policy networks and value networks to guide search, progressively constructing formal proofs.
    • GPT‑f, Lean‑Dojo, etc.: Trained on large-scale formalized theorem and proof corpora, used to automatically generate proofs in systems such as Lean.
  • Mathematical large models and tool augmentation:
    • Minerva, Gödel, etc.: Large models fine-tuned on corpora such as math textbooks, papers, and problem banks, demonstrating stronger performance on proof problems, competition problems, and higher-order reasoning tasks.
    • LLM + SymPy / Mathematica / Lean / Coq: LLMs handle problem parsing and strategy planning, invoking symbolic computation and proof tools for precise operations and verification.
  • Products and applications:
    • "Math teaching assistants / problem-solving assistants" in educational products, providing personalized explanations and multiple solution paths.
    • Research assistance tools: Help researchers construct conjectures, generate proof drafts, search for related theorems and lemmas, accelerating theoretical exploration.
    • Formula derivation and risk model analysis in engineering / finance: Formalize complex models, perform symbolic sensitivity analysis and compliance review.

10.5.1 Automated Theorem Proving and Formal Reasoning

Automated Theorem Proving (ATP) and Interactive Theorem Proving (ITP) are important directions at the intersection of mathematics and computer science. The core task for AI in this domain is to automatically construct or assist in constructing proofs within formal systems, reducing the human burden on low-level details and allowing more focus on high-level ideas.

In formal systems:

• Theorems are encoded as goal types to be constructed, and proofs correspond to constructing a term whose type matches the goal type.
• The proof process consists of a series of tactics or reasoning steps, each advancing under strict logical rules.

AI can play multiple roles in this:

1. Tactic selection and parameter recommendation: Given the current proof state, predict the next tactic to use and its parameters, reducing manual trial-and-error and backtracking.
2. Lemma and theorem retrieval: Retrieve the most relevant lemmas / theorems from large libraries for the current goal, narrowing the search space.
3. End-to-end proof generation: Given a theorem and context, directly generate complete or partial proof scripts, which are then verified for correctness by the prover.

AlphaZero‑style provers, GPT‑f, Lean‑Dojo, and other works, by training policy and value networks or language models on large-scale formalized corpora, have achieved automatic proof of a considerable proportion of theorems in systems such as Lean / Coq. In terms of product direction, this capability is expected to evolve into "formal verification assistants" for software / hardware verification, cryptographic protocol analysis, and high-reliability system design.

10.5.2 Symbolic Computation and Mathematical Problem Solving: LLM + CAS

Compared to theorem proving, symbolic computation and mathematical problem solving are closer to engineering and education scenarios. The goal is: starting from natural language problems, automatically construct symbolic expressions, perform computations, and produce interpretable solution steps.

In this direction, the typical neural–symbolic collaboration workflow is:

1. Problem understanding and abstraction: The LLM parses problems in natural language or images into structured mathematical expressions (equations, constraints, objective functions, etc.).
2. Symbolic expression generation: Translate the abstracted results into CAS code (e.g., SymPy expressions, Mathematica commands).
3. Invoke CAS for execution: Use CAS for precise algebraic operations, differentiation, integration, solving systems of equations, limits, etc.
4. Result interpretation and step generation: The LLM, based on CAS computation results, generates human-readable solution steps and explanations.

This model has several key advantages:

• Correctness of computation is guaranteed by CAS, avoiding LLM "misaligned operations" and accumulated errors on long expressions.
• The LLM provides natural language understanding and expression, lowering the barrier to using CAS and enabling non-expert users to invoke powerful symbolic tools.
• In educational scenarios, the level of detail and style of explanation can be controlled, generating expositions suitable for different learning stages.

In engineering / finance scenarios, this capability can be extended to the formalization and analysis of complex models: automatically extracting model structures from documents and code, constructing symbolic representations, and performing sensitivity analysis, boundary case analysis, and risk identification.

10.6 Scientific Workflow & Lab Automation

The preceding sub-directions mostly focus on "single-point capabilities": predicting a property, generating a structure, proving a theorem. However, in real-world scientific and industrial R&D, what is more critical is how to chain these capabilities into complete workflows and connect them with literature, databases, simulation platforms, and automated experimental equipment. The scientific workflow and lab automation direction aims to build integrated Agent + Tools + Robots systems for scientific scenarios, enabling AI to evolve from "knowing how to compute" to "knowing how to run experiments and conduct research."

This direction connects on one end to paper and patent databases (e.g., PubMed, arXiv), scientific data warehouses, domain knowledge graphs, and simulation platforms, and on the other end to robotic labs, high-throughput screening equipment, and research process management systems. The following unfolds from three perspectives — scenarios, principles, and models.

• Scenarios
  • Scientific literature mining and knowledge base construction: Automatically extract compounds, proteins, materials, reaction conditions, experimental results, and other information from massive papers, building structured knowledge bases and knowledge graphs.
  • Experimental design and Self‑Driving Lab: Under AI-proposed experimental plans, robotic experiment platforms automatically execute formulation, reactions, measurements, and data acquisition, achieving "closed-loop" optimization.
  • Scientific data management and reproducibility assurance: Automatically organize simulation and experimental data, metadata, and code scripts, generating standardized experimental records and reports to improve traceability and reproducibility.
  • Domain "AI experiment assistants": Provide one-stop literature retrieval, protocol design, experimental planning, and result analysis support for pharmaceutical companies, materials companies, and research institutions.
• Principles
  • Literature mining and domain LLMs:
    • Use domain-pretrained models such as SciBERT, BioBERT, and PubMedBERT for named entity recognition, relation extraction, reaction parsing, and experimental condition extraction.
    • On this foundation, train domain LLMs such as Bio‑LM, Chem‑LM, and Materials‑LM to improve understanding and reasoning of specialized terminology, experimental statements, and implicit assumptions.
  • Experimental design and Self‑Driving Lab:
    • Treat the experimental space (formulation, temperature, time, addition order, etc.) as optimization variables, with LLM + RL or Bayesian optimization strategies proposing the next set of experimental conditions.
    • Experimental robots and instruments execute according to the plan, collect data and return it in real time, with models updating parameters and uncertainty estimates, forming an active learning closed loop.
  • Workflow orchestration and agents:
    • Under the Agent & Tool Use framework, unify literature retrieval, code generation, simulation invocation, data analysis, visualization, and report generation tools.
    • Agents, based on task objectives (e.g., "find a high-conductivity electrolyte formulation"), automatically plan task decomposition, tool invocation order, and result integration.
• Models
  • Literature and knowledge mining models:
    • SciBERT, BioBERT, PubMedBERT, etc.: Models pretrained on scientific and biomedical literature for entity / relation extraction, classification, and question answering.
    • Galactica, domain-specialized LLMs: Trained primarily on scientific corpora, supporting review generation, code drafts, experimental design suggestions, etc.
  • Experimental planning and control models:
    • LLM + RL / Bayesian Optimization: Combine domain priors, model uncertainty, and experimental cost to efficiently explore and exploit the experimental space.
    • Agents integrated with Robotic Lab control interfaces: Translate natural language experimental descriptions into structured experimental steps and instrument control commands.
  • Scientific agents and workflow systems:
    • Building on the Agent & Tool Use capabilities from Chapter 7, construct "multi-tool agents" for scientific scenarios: able to retrieve literature, generate code, invoke simulations, process data, create charts, and write report drafts.
  • Products and applications:
    • "AI experiment assistants" and automated experiment benches within pharmaceutical / materials companies: used to accelerate formulation development, process optimization, and candidate screening.
    • Domain scientific search engines and knowledge graphs (Bio / Chem / Materials / Physics Knowledge Graph): Support semantic search, interactive exploration, and knowledge reasoning.
    • Research process management platforms: Integrate experimental planning, data recording, version management, visualization, and automated report generation, improving research team efficiency and result reproducibility.

10.6.1 Scientific Literature Mining and Domain Knowledge Base Construction

The vast majority of scientific knowledge first appears in the form of papers and reports. For AI to truly participate in research, it must be able to "read papers and extract structured knowledge from them." Scientific literature mining and knowledge base construction is precisely about building queryable, reason-able knowledge infrastructure from unstructured text.

In this direction, core tasks include:

• Entity recognition and standardization: Identify entities such as compounds, proteins, materials, reactants, products, experimental equipment, and conditions in literature, and align them with standard databases (e.g., ChEMBL, Uniprot, Materials Project).
• Relation and event extraction: Extract relationships and events such as "who interacts with whom and how" and "what results were produced under what conditions" from text, e.g., reaction equations, formulation–performance correspondences, etc.
• Knowledge graph construction: Organize entities and relations into graph structures, supporting complex queries (e.g., "all reported methods for improving a certain property under certain conditions") and path reasoning.

To achieve these goals, the following are commonly adopted:

• Pretrained models such as SciBERT, BioBERT, and PubMedBERT for NER, RE, and document-level event extraction.
• On this foundation, build domain-specialized LLMs (Bio‑LM, Chem‑LM, Materials‑LM) for more complex question answering, review generation, and knowledge completion.

The constructed domain knowledge bases and knowledge graphs not only provide smarter retrieval and recommendation services for R&D personnel, but also supply data and prior support for subsequent experimental design and materials / drug inverse design.

10.6.2 Self‑Driving Lab and Scientific Workflow Agents: From "Reading Papers" to "Running Experiments"

With literature mining, modeling, and optimization capabilities in place, the next step is to combine these capabilities with automated experimental platforms to build truly Self‑Driving Labs and scientific workflow agents.

In a Self‑Driving Lab, the typical closed-loop workflow is:

1. Goal setting: The researcher provides a macro-level goal (e.g., "improve the conductivity of a certain material under specific conditions") and constraints (cost, safety, process limitations, etc.).
2. Literature and knowledge retrieval: The agent invokes literature retrieval and knowledge graphs to understand existing work and empirical patterns, forming initial hypotheses and an experimental design space.
3. Experimental planning and optimization strategy: Based on LLM + RL / Bayesian optimization strategies, propose the first batch of experimental conditions (formulation, temperature, time, environment, etc.).
4. Robotic execution and data acquisition: The automated experiment bench (Robotic Lab) executes the experiments, collects results in real time, and feeds them back.
5. Model update and next-round design: The surrogate model updates parameters and uncertainty estimates based on new data, then proposes the next round of more informative or promising experimental conditions.

In the broader scientific workflow agent, this closed loop extends to simulation, data analysis, and report generation:

• Agents can automatically generate simulation code or invoke existing simulation tools to pre-evaluate certain experimental conditions;
• During the data analysis phase, automatically complete data cleaning, visualization, and statistical testing;
• At project milestone summaries, generate structured experimental records and report drafts with charts and references.

In terms of product form, such systems are typically delivered as platforms: providing a unified interface and API, connecting to literature databases, simulation engines, and experimental equipment, allowing scientists and engineers to set goals at a high level using natural language and visual interfaces, with the remaining steps automatically orchestrated and executed by agents + toolchains.

From this sub-direction onward, AI's role in science truly shifts from "offline analysis tool" to "online research collaborator": not only able to read papers, write code, and compute models, but also to work alongside robots to complete real experiments and discoveries.