Aug 2026· ENERGY & ENVIRONMENTAL MATERIALS· 0 citations· 99 references
TL;DR
A review of physically constrained, multimodal and closed‐loop workflows as the clearest route from generative crystal models to experimentally actionable candidates and defines the inverse‐design problem and main generation tasks.
Abstract
Discovery of inorganic crystals remains limited by trial‐and‐error experimentation and by the cost of exploring coupled chemical, compositional and structural variables. Generative inverse design addresses this limitation by sampling candidate crystal structures from target properties rather than screening only predefined libraries. Its practical value, however, depends on three technical requirements: a representation that preserves crystallographic information, generative architectures that respect physical constraints, and evaluation protocols that distinguish valid structures from stable and synthesizable candidates. This review examines these requirements for inorganic crystal discovery. We first define the inverse‐design problem and the main generation tasks, then compare variational autoencoders, generative adversarial networks, diffusion models, flow‐based models, large language model pipelines and hybrid frameworks. The comparison shows architecture‐dependent trade‐offs. Variational autoencoders provide continuous latent spaces but often require post‐generation validation; Generative adversarial networks improve distributional sampling but remain sensitive to training instability and mode collapse; diffusion models offer strong conditional control at the cost of iterative sampling and screening; flow‐based methods provide efficient, symmetry‐aware generation; Large language model‐based approaches require standardized and symmetry‐consistent tokens; and hybrid systems assign composition, symmetry, lattice and coordinate generation to separate modules. We then discuss battery and energy materials as application testbeds and review representation choices, data resources and benchmarks. The review identifies physically constrained, multimodal and closed‐loop workflows as the clearest route from generative crystal models to experimentally actionable candidates.
A reinforcement learning framework that guides latent denoising diffusion models in finding diverse and novel, yet thermodynamically viable, crystalline compounds and demonstrates enhanced property-guided design that preserves chemical validity while targeting desired functional properties.
Hyunsoo Park, Aron Walsh· Nature Machine Intelligence· 27 citations· ⚡3
Generative artificial intelligence is rapidly transforming materials design by enabling de novo exploration of immense chemical spaces. Yet a large proportion of AI-generated compositions remain implausible, violating established chemical principles, which limits the reliability and interpretability of generative materials design. Here, we introduce a chemical validity operator that recasts heuristic chemical rules as a configurable algorithmic prior for evaluating and guiding generative materials discovery. Built on the open-source SMACT package, a data-informed oxidation-state model exposes tunable thresholds, allowing users to interpolate continuously between permissive and conservative chemical constraints, while supporting both exploratory and conservative materials-design workflows. Benchmarking six state-of-the-art generative models for inorganic crystals shows that most reproduce stoichiometry but under-represent realistic oxidation-state combinations, and that filtering removes compositions reliant on rarely observed oxidation states while preserving low-energy compounds near the convex hull. Beyond screening, the same operator can also serve as a reinforcement-learning reward, steering a latent diffusion model towards chemically grounded compositions. By encoding chemical heuristics and observations, this work establishes a foundation for oxidation-state-aware generative models.
Kinga O. Mastej, Panyalak Detrattanawichai, Hyunsoo Park et al.· 0 citations
Atomistic simulations have become indispensable tools for understanding glass structure, dynamics, and properties, yet persistent challenges limit their predictive power. This perspective examines three interconnected issues, namely, glass formation procedures, interatomic potential development, and machine learning (ML) applications, which emerged from the 5th International Workshop on Challenges of Atomistic Simulations of Glasses and Amorphous Materials. We identify convergent community priorities for standardized validation protocols and curated benchmark datasets with complete metadata. A systematic pathway forward is provided by a hierarchical validation framework for assessing the structural fidelity, property prediction, and behavioral realism of simulation techniques. Looking ahead, transformative advances are promised by the fusion of classical techniques with ML‐based approaches. Examples include the integration of swap Monte Carlo with ML potentials, leveraging foundation models through transfer or active learning, and fine‐tuning ML potentials with experimental data. Progress depends on community awareness and efforts committed to validating models, creating reproducible protocols, and sustained data sharing.
N. M. A. Krishnan, A. Pedone, Xiaonan Lu et al.· Journal of The American Cera...· 0 citations
Closed-loop materials discovery iterates between proposing candidate structures and evaluating their properties, and property evaluation dominates the cost. In the generative variant, a learned prior proposes candidate crystals and a property oracle scores them; we ask whether a cheap probabilistic surrogate can triage the generator's output, and what such a surrogate must do well. Across three architecturally distinct pretrained diffusion priors (MatterGen, CrystalFlow, ADiT) and two targets (room-temperature heat capacity and bulk modulus), we insert a Gaussian process acquisition gate between structure generation and the oracle in an RL-steered generative workflow. The gate matches or exceeds ungated fine-tuning of the generative model while capping oracle calls at a fixed per-cycle budget. Budget-matched ablations isolate the mechanism. At an identical four-call budget, ranking-based selection outperforms arbitrary selection, confirming that the gain comes from the surrogate's choice; the gate comes within $\sim$9\% of exhaustive oracle spending at roughly one-fifth of the calls. A density-functional-theory check of the bulk-modulus discoveries confirms the learned oracle to within 2.5\% on average and the surrogate's ranking of the generated structures at Spearman $\rho = 0.94$. A cross-factorial benchmark of surrogate performance spanning mechanical, electronic, and vibrational properties identifies pretrained ORB embeddings with a Gaussian process as the most reliable combination, which we adopt as the building blocks of the proposed workflow. The complete pipeline is released as open-source software.
Sk Md Ahnaf Akif Alvi, Jan Janssen, Danny Perez et al.· arXiv.org· 0 citations
The vast chemical design space and complex, interdependent design variables make catalyst discovery for targeted properties highly labor- and resource-intensive. Although generative models have emerged as a promising solution, existing approaches are generally limited to single-property conditioning or narrow chemical spaces. Here, we present Catalyst Diffusion Transformer (CatDiT), a unified framework for inverse catalyst design that generates valid and novel structures ranging from intermetallic alloys to oxide surfaces. By learning compressed latent representations, CatDiT enables efficient training and rapid sampling while supporting simultaneous conditioning on adsorbate type, binding energy, and catalyst class. The model provides reliable control of discrete properties and directional control of continuous properties, enriching candidate pools for reaction-specific catalyst discovery. As a representative application, multi-conditional generation for the nitrogen reduction reaction (NRR) yields 28 density functional theory (DFT)-relaxed alloy candidates that satisfy the target activity window and lie above the pure-metal *N-*H scaling line, corresponding to a ~1.5-fold enrichment over the source distribution. These results establish CatDiT as a practical and scalable approach for property-directed catalyst inverse design and targeted catalyst generation.
Hayoung Doo, Dong Hyeon Mok, S. Back et al.· 0 citations
Generative AI and high-throughput DFT pipelines propose millions of inorganic crystal structures, but lack a calibrated reference frame against experimentally realized chemistry. Here we embed 167,500 Inorganic Crystal Structure Database entries in a continuous structural-similarity space, partition it into graph communities, and replay them in time. Experimental discovery shows strong structural memory: 82.9% of new formulas enter pre-existing communities; new-community formation falls from 40.2% (1930s) to 2.6% (2010s). The communities are chemically meaningful, positively identifying nine textbook field-defining renaissances, including cuprates, colossal-magnetoresistance manganites, MAX phases, and Li-ion battery cathodes. Projecting GNoME, MatterGen-public, Materials Project, JARVIS-DFT, and Alexandria-PBE into frozen historical maps yields a cutoff-robust ordering: held-out ICSD>MatterGen>{GNoME ~ MP-theoretical}>JARVIS>Alexandria. Structural departure from experimental basins is not specific to generative AI but general across the tested computed sets. Combining structural proximity with reduced-formula precedent defines a historical synthesizability prior for triaging computed materials.
D. Nguyen, Karen Cao, Brian Chu et al.· 0 citations