How is AI accelerating Materials Discovery—and why are GFlowNets a core ingredient ?

Over the past two years, AI has made significant strides in materials science, with major actors like Google DeepMind, Microsoft, and Meta releasing high-impact research that changes how we explore the chemical space.

 ​

In 2023, researchers at Google DeepMind managed to generate millions of theoretically stable materials with their model called GNoME [1] (published in Nature). They achieved it by training a neural network to predict the stability of new inorganic crystals created by (randomly) mutating known ones. This predictive model was then refined in a process called active learning, where the model selected the most informative candidates and evaluated them using quantum simulations to further improve the accuracy of their predictions . But while GNoME used AI purely to estimate the stability of these new candidate materials, models like MatterGen [2] from Microsoft Research (published in Nature earlier this year) go one step further: rather than exploring randomly around inorganic crystals known to chemists, they leverage AI to generate, from scratch, new materials that match target properties such as a specific band gap.

 ​

These advances point to a powerful idea: AI can perform an intelligent first identification of candidates in the chemical space, concentrating on promising regions instead of searching completely randomly. This is essential because each evaluation has a cost. Whether it’s a quantum simulation like density functional theory (DFT) or a real-world lab experiment, every evaluation demands significant time and resources. Therefore, success should not be measured only in terms of the quantity of materials being screened, but in the way they are prioritized for deeper investigation.

 ​

These ML models are usually only as good as the data on which they are trained. That’s why foundational datasets like LeMat-Bulk & LeMat-Traj [3], which we recently released in collaboration with HuggingFace, are so important. By unifying and deduplicating millions of quantum chemistry results from databases like Materials Project, OQMD, and Alexandria, LeMaterial series provides a clean and consistent foundation for training the next generation of generative AIs.

 ​

However, chemists and material scientists often still perceive AI as a black box, undoubtedly powerful but opaque. This does not foster trust and usage, especially as purely data-driven models can sometimes generate candidates that violate basic chemical intuitions, which experts would spot immediately. That’s why at Entalpic, we focus on building a generative AI that is transparent, controllable, and grounded in chemistry. Our objective is to empower domain experts, not replace them.