AI-driven materials discovery combines large pretrained atomistic models, generative inverse design and robotic closed-loop labs to propose and screen candidate materials at scale; the 2026 literature shows candidate generation has become cheap and abundant while synthesis, decision-making and experimental verification are now the binding constraints.
Summary
Materials discovery has historically been forward-looking and slow: pick a composition, compute or measure its properties, iterate. The stack described across the 2026 arXiv literature inverts this. At the bottom sit large curated structure databases, now themselves expanded by AI rather than by density functional theory (DFT) alone. Above them sit pretrained models: molecular foundation models such as MIST, trained with a tokenizer designed to capture nuclear, electronic and geometric information and reported at up to an order of magnitude more parameters and data than prior work, and machine learning interatomic potentials (MLIPs) built on graph neural networks that approximate DFT accuracy at far lower cost.
On top of that sit property predictors and generators. Predictors range from black-box graph models to composition-weighted symbolic regression that returns explicit analytical expressions while remaining competitive on MatBench benchmarks, and multimodal models that fuse structure with spectroscopy, microscopy and processing context. Generators perform inverse design: state a target property, get structures back. Recent variants inject continuous property representations directly into a transformer’s attention to condition crystal generation, use symmetry-constrained generation pinned to Wyckoff positions to reach geometries outside the known kagome, Lieb and pyrochlore catalogue, or skip bespoke generators entirely by running gradient optimisation through off-the-shelf pretrained predictors.
The layer receiving most attention in 2026 is the one that closes the loop. Agentic frameworks orchestrate atomic models for numerical work and language models for semantic judgement, because proposing millions of candidates is no longer the hard part; choosing which to make is. Reliability-gated pipelines admit first-principles results as training feedback only when provenance and quality checks pass. Self-driving laboratories (SDLs) pair AI proposal with robotic execution, mediated by software that decouples algorithms from heterogeneous hardware.
The parameters that decide the field are now well identified by its own practitioners: the synthesizability gap between a computed structure and an executable recipe; out-of-distribution generalisation of pretrained potentials when applied to chemically novel targets, and the DFT labelling cost of fixing it; data provenance and uncertainty quantification; cross-institutional composition of models, data and instruments; and the absence of environmental constraints inside the design loop rather than downstream of it.
Evidence base
- 2026-05-05: An agentic framework combining a 1-billion-parameter large atomic model with LLM reasoning rediscovered 66 experimentally verified superconductors absent from the standard SuperCon3D database, then screened 2.4 million equilibrium crystals to identify 68,000 candidates.
- 2026-05-04: A perspective argues the structure-centric AI paradigm is stalling at the synthesizability gap despite delivering thousands of candidate structures, and proposes treating executable synthesis protocols as primary design variables.
- 2026-06-17: Predictor-guided gradient optimisation outperformed generative and Bayesian baselines on perovskite design across band gap, formation energy and tolerance factor under two hard constraints, with DFT-assessed band-gap targeting competitive with a leading generative model using predictors trained on roughly one-tenth of the data.
- 2026-06-16: Symmetry-constrained generation beyond the known kagome, Lieb and pyrochlore motif catalogue produced 9,352 crystal candidates surviving stability and flatness screening, with flatness confirmed by high-throughput full DFT.
- 2026-06-16: A reliability-gated first-principles feedback workflow processed 516 2DMatPedia-derived candidates into 280 QC-passed materials and 573 retained carrier-direction channels, splitting them into separate feedback objects for the generative model and the property model.
- 2026-06-16: An open-source closed-loop platform reports twelve AI algorithms behind a unified Python interface and six deployed self-driving laboratory implementations, using CSV file exchange to decouple algorithms from heterogeneous robotic hardware.
Open questions
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.