Accurate prediction of synthesizability and precursors of 3D crystal structures via large language models

Zhilong Song, Shuaihua Lu, Minggang Ju, Qionghua Zhou, Jinlan Wang

Accessing the synthesizability of crystal structures is crucial for transforming theoretical materials into real-world applications. Nevertheless, there is a significant gap between actual synthesizability and thermodynamic or kinetic stability commonly used to screen synthesizable structures. Herein, we develop the Crystal Synthesis Large Language Models (CSLLM) framework, which utilizes three specialized LLMs to predict the synthesizability of arbitrary 3D crystal structures, possible synthetic methods, and suitable precursors, respectively. We construct a comprehensive dataset including synthesizable/non-synthesizable crystal structures and develop an efficient text representation for crystal structures to fine-tune LLMs. Our Synthesizability LLM achieves state-of-the-art accuracy (98.6%), significantly outperforming traditional synthesizability screening based on thermodynamic and kinetic stability. Its outstanding generalization ability is further demonstrated in experimental structures with complexity considerably exceeding that of the training data. Furthermore, both the Method and Precursor LLMs exceed 90% accuracy in classifying possible synthetic methods and identifying solid-state synthetic precursors for common binary and ternary compounds, respectively. Leveraging CSLLM, tens of thousands of synthesizable theoretical structures are successfully identified, with their 23 key properties predicted using accurate graph neural network models.

Inverse design of promising electrocatalysts for CO₂ reduction via generative models and bird swarm algorithm

Zhilong Song, Linfeng Fan, Shuaihua Lu, Chongyi Ling, Qionghua Zhou, Jinlan Wang

Directly generating material structures with optimal properties is a long-standing goal in material design. Traditional generative models often struggle to efficiently explore the global chemical space, limiting their utility to localized space. Here, we present a framework named Material Generation with Efficient Global Chemical Space Search (MAGECS) that addresses this challenge by integrating the bird swarm algorithm and supervised graph neural networks, enabling effective navigation of generative models in the immense chemical space towards materials with target properties. Applied to the design of alloy electrocatalysts for CO₂ reduction (CO₂RR), MAGECS generates over 250,000 structures, achieving a 2.5-fold increase in high-activity structures (35%) compared to random generation. Five predicted alloys— CuAl, AlPd, Sn₂Pd₅, Sn₉Pd₇, and CuAlSe₂ are synthesized and characterized, with two showing around 90% Faraday efficiency for CO₂RR. This work highlights the potential of MAGECS to revolutionize functional material development, paving the way for fully automated, artificial intelligence-driven material design.

Stability and Symmetry-Assured Crystal Structure Generation for Inverse Design of Photocatalysts in Water Splitting

Zhilong Song, Chongyi Ling, Qiang Li, Qionghua Zhou, Jinlan Wang

Generative models are revolutionizing materials discovery by enabling inverse design-direct generation of structures from desired properties. However, existing approaches often struggle to ensure inherent stability and symmetry while precisely generating structures with target compositions, space groups, and lattices without fine-tuning. Here, we present SSAGEN (Stability and Symmetry-Assured GENerative framework), which overcomes these limitations by decoupling structure generation into two distinct stages: crystal information (lattice, composition, and space group) generation and coordinate optimization. SSAGEN first generates diverse yet physically plausible crystal information, then derives stable and metastable atomic positions through universal machine learning potentials, combined global and local optimization with symmetry and Wyckoff position constraints, and dynamically refined search spaces. Compared to prior generative models such as CDVAE, SSAGEN improves the thermodynamic and kinetic stability of generated structures by 148% and 180%, respectively, while inherently satisfying target compositions, space groups, and lattices. Applied to photocatalytic water splitting (PWS), SSAGEN generates 200,000 structures-81.2% novel-with 3,318 meeting all stability and band gap criteria. Density functional theory (DFT) validation confirms 95.6% structures satisfy PWS requirements, with 24 optimal candidates identified through comprehensive screening based on electronic structure, thermodynamic, kinetic, and aqueous stability criteria. SSAGEN not only precisely generates materials with desired crystal information but also ensures inherent stability and symmetry, establishing a new paradigm for targeted inverse design of functional materials.

LLM-Feynman: Leveraging Large Language Models for Universal Scientific Formula and Theory Discovery

Zhilong Song, Minggang Ju, Chunjin Ren, Qiang Li, Chongyi Li, Qionghua Zhou, Jinlan Wang

Distilling underlying principles from data has historically driven scientific breakthroughs. However, conventional data‐driven machine learning often produces complex models that lack interpretability and generalization due to insufficient domain expertise. Here, we present LLM-Feynman, a novel agent that leverages large language models (LLMs) alongside systematic optimization to derive concise, interpretable formulas from data and domain knowledge. Our method integrates automated feature engineering, LLM-guided symbolic regression with self-evaluation, and Monte Carlo tree search to reduce LLM hallucination, thereby enhancing formula discovery. The embedding of domain knowledge simplifies the formula, while self-evaluation based on this knowledge further minimizes prediction errors, surpassing conventional symbolic regression in accuracy and interpretability. Validation on datasets from Feynman physics lectures confirms that LLM-Feynman can rediscover over 90% real physical formulas. Moreover, when applied to four key materials science tasks – from classifying the synthesizability of 2D and perovskite structures to predicting ionic conductivity in lithium solid-state electrolytes and GW bandgaps in 2D materials – LLM-Feynman consistently yields interpretable formula with accuracy exceeding 90% and R2 values above 0.8. By transcending mere data fitting through the integration of deep domain knowledge, LLM-Feynman establishes a new paradigm for the automated discovery of generalizable scientific formula and theory across disciplines.

T2MAT (text-to-materials): A universal agent for generating material structures with goal properties from a single sentence

Zhilong Song, Shuaihua Lu, Qionghua Zhou, Jinlan Wang

Artificial Intelligence-Generated Content (AIGC)—content autonomously produced by AI systems without human intervention—has significantly boosted efficiency across various fields. However, AIGC in material science faces challenges in efficiently discovering novel materials that surpass existing databases, while ensuring the invariance and stability of crystal structures. To address these challenges, we develop T2MAT (text-to-material), an end-to-end agent that transforms user-input text into the inverse design of novel material structures with target properties beyond existing database, enabled by comprehensive exploration of chemical space and fully automated first-principles validation. Furthermore, we propose CGTNet (Crystal Graph Transformer NETwork), a graph neural network specifically designed to capture long-range interactions, which dramatically improves the accuracy and data efficiency of property predictions and thereby strengthens the reliability of inverse design. Through these contributions, T2MAT reduces the reliance on human expertise and accelerates the discovery of high-performance functional materials, paving the way for truly autonomous material design.

Simple descriptor derived from symbolic regression accelerating the discovery of new perovskite catalysts

Baicheng Weng#, Zhilong Song#, Rilong Zhu, Qingyu Yan, Qingde Sun, Corey G Grice, Yanfa Yan, Wan-Jian Yin

Symbolic regression (SR) is an approach of interpretable machine learning for building mathematical formulas that best fit certain datasets. In this work, SR is used to guide the design of new oxide perovskite catalysts with improved oxygen evolution reaction (OER) activities. A simple descriptor, μ/t, where μ and t are the octahedral and tolerance factors, respectively, is identified, which accelerates the discovery of a series of new oxide perovskite catalysts with improved OER activity. We successfully synthesise five new oxide perovskites and characterise their OER activities. Remarkably, four of them, Cs_0.4La_0.6Mn_0.25Co_0.75O₃, Cs_0.3La_0.7NiO₃, SrNi_0.75Co_0.25O₃, and Sr_0.25Ba_0.75NiO₃, are among the oxide perovskite catalysts with the highest intrinsic activities. Our results demonstrate the potential of SR for accelerating the data-driven design and discovery of new materials with improved properties.

Distilling universal activity descriptors for perovskite catalysts from multiple data sources via multi-task symbolic regression

Zhilong Song, Xiao Wang, Fangting Liu, Qionghua Zhou, Wan-Jian Yin, Hao Wu, Weiqiao Deng, Jinlan Wang

Developing activity descriptors via data-driven machine learning (ML) methods can speed up the design of highly active and low-cost electrocatalysts. Despite the fact that a large amount of activity data for electrocatalysts is stored in the literature, data from different publications are not comparable due to different experimental or computational conditions. In this work, an interpretable ML method, multi-task symbolic regression, was adopted to learn from data in multiple experiments. A universal activity descriptor to evaluate the oxygen evolution reaction (OER) performance of oxide perovskites free of calculations or experiments was constructed and reached high accuracy and generalization ability. Utilizing this descriptor with Bayesian-optimized parameters, a series of compelling double perovskites with excellent OER activity were predicted and further evaluated using first-principles calculations. Finally, the two ML-predicted nickel-based perovskites with the best OER activity were successfully synthesized and characterized experimentally. This work opens a new way to extend machine-learning material design by utilizing multiple data sources.

Adaptive Design of Alloys for CO₂ Activation and Methanation via Reinforcement Learning Monte Carlo Tree Search Algorithm

Zhilong Song, Qionghua Zhou, Shuaihua Lu, Sae Dieb, Chongyi Ling, Jinlan Wang

Data-driven machine learning (ML) has earned remarkable achievements in accelerating materials design, while it heavily relies on high-quality data acquisition. In this work, we develop an adaptive design framework for searching for optimal materials starting from zero data and with as few DFT calculations as possible. This framework integrates automatic density functional theory (DFT) calculations with an improved Monte Carlo tree search via reinforcement learning algorithm (MCTS-PG). As a successful example, we apply it to rapidly identify the desired alloy catalysts for CO₂ activation and methanation within 200 MCTS-PG steps. To this end, seven alloy surfaces with high theoretical activity and selectivity for CO₂ methanation are screened out and further validated by comprehensive free energy calculations. Our adaptive design framework enables the fast computational exploration of materials with desired properties via minimal DFT calculations.