Materials Discovery Patents

AI materials discovery is a rapidly emerging patent space — combining computational chemistry; machine learning; and high-throughput experimentation — with major activity from technology companies; chemical companies; and national laboratories: MAJOR AI MATERIALS DISCOVERY PATENT HOLDERS: GOOGLE DEEPMIND / GOOGLE: 2,000+ materials science and AI patents; specific GNoME (Graph Networks for Materials Exploration: specific graph neural network crystal structure stability prediction; specific specific trained on 48,000+ experimentally known stable crystals from specific specific ICSD + specific specific MP database; specific specific predicted 2.2 million new stable crystal structures including specific specific 380,000 stable materials beyond previous knowledge as of 2023 publication; specific specific formation energy prediction within specific specific 10 meV/atom DFT accuracy via specific specific message passing architecture); specific AlphaFold2 protein structure prediction by analogy (specific specific invariant point attention IPA + specific specific Evoformer for specific specific atom coordinate prediction from specific specific amino acid sequence — directly enabling protein engineering materials discovery); IBM (3,000+; specific RXN for Chemistry reaction prediction; specific specific chemistry-aware transformer for specific specific reaction outcome + specific specific retrosynthesis; specific specific molecular property prediction); MICROSOFT (AZURE QUANTUM; PREVIOUSLY CAMBRIDGE): specific Azure Quantum Elements (specific specific density functional theory DFT computation at scale; specific specific high-throughput screening workflow for specific specific battery electrolyte + catalyst candidate materials); BASF; SOLVAY; DOW: 10,000+ each; specific computational materials design for specific polymer + specific catalyst + specific coating; specific high-throughput synthesis + characterization pipeline; NATIONAL LABORATORIES: DOE: ORNL; ANL; LBNL; NREL; PNNL hold extensive materials IP; specific DFT-driven battery electrode material prediction; specific specific ML interatomic potentials for specific specific molecular dynamics at DFT accuracy; ACADEMIA TO STARTUP PIPELINE: MIT; Stanford; CMU; UC Berkeley materials discovery research generates significant spinout IP.

Graph neural networks for property prediction; generative AI for novel materials design; and crystal structure prediction from first principles are the three most active technical areas in AI materials discovery IP — each with distinct patentability considerations: GRAPH NEURAL NETWORK MATERIALS PROPERTY PREDICTION PATENTS: SPECIFIC PATENTABLE INNOVATIONS: specific message passing neural network MPNN for specific crystal structure (specific specific atom-bond graph construction from specific specific crystallographic input CIF + specific specific periodic boundary conditions for specific specific property prediction: specific specific formation energy within specific specific X meV/atom vs. DFT; specific specific band gap within specific specific X eV vs. experiment; specific specific elastic modulus Gpa; specific specific specific application domain); specific equivariant graph neural network architecture for specific materials property (specific specific SE(3)-equivariant or specific specific E(3)-equivariant representation with specific specific spherical harmonics basis for specific specific rotationally invariant property prediction from specific specific atom coordinate + specific specific element feature); specific multi-fidelity materials property prediction (specific specific transfer learning from specific specific low-fidelity DFT-PBE to specific specific high-fidelity HSE06 + specific specific experiment with specific specific measured accuracy improvement on specific specific materials class); GENERATIVE AI MATERIALS DESIGN PATENTS: DIFFUSION MODELS FOR CRYSTAL STRUCTURE GENERATION: MIT; STANFORD; CARNEGIE MELLON (CDVAE; DIFFCSP): specific periodic crystal structure generative model (specific specific score-based diffusion over specific specific atom coordinates + specific specific lattice parameters with specific specific periodic boundary condition for specific specific crystal stability — measured by specific specific DFT formation energy on generated candidates); specific conditional crystal generation (specific specific property-conditioned generation: specific specific target band gap or specific specific ionic conductivity + specific specific space group constraint for specific specific inverse design of specific specific novel crystal for specific specific application); GENERATIVE CHEMISTRY: INSILICO MEDICINE; RECURSION; EXSCIENTIA: specific specific molecule generation by specific specific transformer or specific specific VAE or specific specific flow-based model with specific specific synthesizability filter and specific specific property objective; CRYSTAL STRUCTURE PREDICTION PATENTS: CCDC; M. SCHEFFLER GROUP; HARVARD: specific ab initio crystal structure prediction (specific specific random structure sampling + specific specific evolutionary algorithm AIRSS or specific specific CALYPSO + specific specific DFT geometry optimization for specific specific global minimum energy structure without specific specific prior experimental knowledge).

High-throughput computational screening; machine learning interatomic potentials (MLIPs); and large-scale materials databases are three foundational IP areas in computational materials discovery — enabling discovery of novel materials for batteries; catalysis; semiconductors; and structural applications: HIGH-THROUGHPUT MATERIALS SCREENING PATENTS: SPECIFIC HTS PIPELINE PATENTS: NREL; ARGONNE; LBNL (MATERIALS PROJECT): specific automated DFT workflow (specific specific workflow management — specific specific AiiDA or specific specific FireWorks + specific specific VASP/Quantum ESPRESSO DFT code for specific specific parallel structure relaxation + specific specific property extraction with specific specific provenance tracking for specific specific reproducibility at specific specific HPC scale); specific multi-property screening funnel (specific specific primary filter: specific specific formation energy stability threshold → specific specific secondary filter: specific specific target property range → specific specific tertiary filter: specific specific synthesizability criterion for specific specific efficient candidate narrowing from specific specific 10^6 candidates to specific specific experimental priority list); MACHINE LEARNING INTERATOMIC POTENTIALS (MLIP) PATENTS: DEEPMIND (DEEPMIND JAX MD; ORCA); MICROSOFT (MATTERSIM; MACE-MP-0); META AI (FAIR; OMAT24); ORION: specific universal MLIP (specific specific multi-element interatomic potential trained on specific specific large DFT dataset: specific specific Materials Project + specific specific Alexandria + specific specific AFLOW + specific specific OMAT24 350M structures for specific specific near-DFT-accuracy MD simulation at specific specific 10^6x DFT speed for specific specific long-timescale + specific specific large-system dynamics); specific MLIP for specific application domain (specific specific solid electrolyte Li-ion conductor MLIP trained on specific specific DFT NVT trajectory dataset for specific specific Li diffusivity prediction as specific specific function of temperature + specific specific composition with specific specific validation against specific specific experimental AC impedance); MATERIALS DATABASES PATENTS/IP: MATERIALS PROJECT (LBNL); ICSD; CCDC; SPRINGER MATERIALS: databases are generally not patentable (data ≠ patentable process); BUT: specific specific data curation workflow; specific specific property calculation pipeline; specific specific API design = patentable if sufficiently inventive; ROBOTICS-ACCELERATED DISCOVERY: CARNEGIE MELLON (EMERALD); MIT (MULTI ROBOT); UNIVERSITY OF LIVERPOOL; ACCELERATE: specific autonomous experiment platform (specific specific robotic liquid handler for specific specific synthesis + specific specific characterization instrument (XRD+UV-Vis+EIS) for specific specific closed-loop Bayesian optimization: specific specific BO acquisition function selecting next experiment from specific specific posterior GP surrogate to maximize specific specific measured property improvement per experiment count; specific specific measured discovery rate: specific specific X novel candidates in specific specific Y days vs. specific specific manual approach).

AI materials discovery startups — operating at the intersection of software; chemistry; and advanced manufacturing — face a complex IP landscape where both the algorithms and the materials themselves may be patentable, creating opportunities for layered IP protection: AI MATERIALS DISCOVERY STARTUP IP STRATEGY: UNDERSTAND THE DUAL IP OPPORTUNITY: AI MATERIALS DISCOVERY STARTUPS CAN PATENT TWO LAYERS: (1) THE AI/ML METHOD: the prediction/generation algorithm + architecture + training dataset combination; (2) THE DISCOVERED MATERIALS: the novel crystal structures; alloys; polymers; or molecules with specific composition + structure + measured properties; LAYER 2 (SPECIFIC MATERIAL) IS OFTEN STRONGER AND LONGER-LIVED — algorithms become prior art quickly; specific novel material compositions last 20 years as composition-of-matter claims; COMPOSITION OF MATTER IS HIGHLY PATENT-ELIGIBLE: specific novel crystal phase with specific measured property; specific novel alloy composition + microstructure; specific specific polymer backbone with specific specific glass transition + specific specific modulus = strong § 101 eligibility (physical matter = not abstract); § 101 RISK FOR AI METHODS: pure prediction algorithm without physical application = abstract idea risk; SURVIVAL: claim the entire system (specific hardware + specific specific algorithm + specific specific measured discovery rate on specific specific benchmark); or claim the discovered material directly; WHEN TO PATENT IN AI MATERIALS DISCOVERY: SPECIFIC NOVEL AI ARCHITECTURE FOR MATERIALS: specific novel GNN architecture for specific materials class with specific measured DFT-accuracy benchmark improvement (MAE formation energy; bandgap MAE; elastic tensor accuracy) vs. prior GNN; specific novel generative model for specific crystal structure with specific measured stability rate (DFT-computed formation energy hull distance) and specific specific measured novelty rate (Tanimoto similarity); SPECIFIC NOVEL DISCOVERED MATERIAL: specific novel crystal structure with specific space group + composition + unit cell parameters + specific measured property (specific band gap; specific ionic conductivity; specific mechanical strength; specific thermal conductivity) measured by specific experimental technique (XRD+STEM; AC impedance; tensile testing; TDTR thermal measurement); specific novel polymer backbone + architecture with specific measured Tg + modulus + permeability; SPECIFIC NOVEL ROBOTIC DISCOVERY PLATFORM: specific novel closed-loop system (specific specific robotic + specific specific BO algorithm + specific specific measured discovery throughput); TRADE SECRETS: specific trained MLIP + GNN model weights for specific materials class; specific specific active learning curriculum trained on specific specific proprietary DFT dataset; specific specific synthesis protocol for specific specific discovered material class; CRITICAL DISTINCTION — SOFTWARE vs. MATERIAL CLAIMS: file both: method claims (AI discovery) + composition claims (specific discovered material with specific measured properties); composition claims survive algorithm commoditization; KEY FTO: Google DeepMind GNoME crystal stability; Microsoft MatterSim universal MLIP; Meta OAMat24 MLIP; MIT/CMU diffusion crystal generator; CCDC CSD database IP; BASF/Dow/Solvay computational chemistry methods; national lab CRADA IP (ORNL; ANL; LBNL; NREL).