Fueling Generative AI with Atomic-Scale Data

AI models that can generate new materials with desirable properties are an exciting frontier in science. Companies and academic groups—including DeepMind’s GNome and Microsoft’s MatterGen—are already exploring this space. These models hold enormous potential, but they face one critical challenge: they require vast amounts of structural and chemical property training data. While some of this data can be measured experimentally, much of it—especially at the atomic scale—is extremely difficult or even impossible to obtain directly.

At their core, material properties emerge from how molecules interact and arrange themselves. These interactions are governed by the laws of quantum mechanics and statistical physics, meaning that all material behavior can, in principle, be derived from physics-based simulations. Such simulations provide an essentially unlimited source of atomic-scale data, making them a powerful foundation for training generative AI models. Freecurve technology is designed to extract this information efficiently, preserving both accuracy and atomic-level detail. Without this fidelity, AI models are trained on partial and imprecise data. As a result they fail to predictively capture physical reality.

Biology provides a useful contrast. Proteins are largely determined by their amino acid sequence, shaped by evolutionary pressures. This made it possible for breakthrough models like AlphaFold to succeed, drawing on more than 200,000 experimentally determined structures and billions of DNA sequences. Materials science, however, faces a much tougher challenge: materials are far more chemically diverse than proteins and often lack stable structures that can be resolved experimentally through methods like X-ray diffraction. This makes simulation-derived data not just helpful, but essential.

Freecurve addresses this by generating from first principles reliable molecular ensembles using its ArrowNN interatomic potentials. These ensembles capture structural and atomic-scale property data that can be used to train Materials Generative AI models. Given the vastness of chemical space, the development process proceeds in stages, each targeting chemical subspaces relevant to specific industrial applications.

For example, in optimizing battery cathode materials based on metal oxides, the workflow could look like this:

1. Structure generation: Create tens of thousands of candidate structures for small metal oxides interacting with cations (Li⁺, Na⁺, Mg²⁺, etc.) and solvent molecules.

   
   

2. Quantum chemical calculations: Perform high-level quantum chemical computations on the generated structures.

   
   

3. ML potential training: Train highly predictive ArrowNN machine-learned interatomic potentials for metal oxides using the QC data.

   
   

4. Ensemble sampling: Use ArrowNN-based molecular dynamics to generate diverse ensembles of metal oxide structures (different molecular arrangements). Compute atomic-scale properties such as solvation energies and ion-binding free energies.

   
   

5. Generative model refinement: Retrain the Materials Generative AI Model on the simulated ensembles and properties obtained with ArrowNN.

   
   

6. Property-driven design: Apply the fine-tuned generative model to predict compositions and structures of metal oxides with desired target properties.

   
   

7. Iterative improvement: Evaluate the predicted structures with ArrowNN simulations, compute their properties, and incorporate the results into the training set for further refinement of the generative AI model.

   
   

This loop—simulation, training, refinement, and validation—creates a robust, physics-grounded pathway for accelerating materials discovery. By providing atomic-level data that experiments cannot capture, Freecurve enables generative AI to reach its full potential in designing materials for energy, defense, electronics, and beyond.