Generative adversarial networks (GAN) based efficient sampling of chemical space for inverse design of inorganic materials

A major challenge in materials design is how to efficiently search the vast chemical design space to find the materials with desired properties. One effective strategy is to develop sampling algorithms that can exploit both explicit chemical knowledge and implicit composition rules embodied in the large materials database. Here, we propose a generative machine learning model (MatGAN) based on a generative adversarial network (GAN) for efficient generation of new hypothetical inorganic materials. Trained with materials from the ICSD database, our GAN model can generate hypothetical materials not existing in the training dataset, reaching a novelty of 92.53% when generating 2 million samples. The percentage of chemically valid (charge neutral and electronegativity balanced) samples out of all generated ones reaches 84.5% by our GAN when trained with materials from ICSD even though no such chemical rules are explicitly enforced in our GAN model, indicating its capability to learn implicit chemical composition rules. Our algorithm could be used to speed up inverse design or computational screening of inorganic materials.

💡 Research Summary

The paper tackles the long‑standing challenge of efficiently exploring the astronomically large inorganic chemical space for materials with target properties. Traditional approaches rely on exhaustive experimental synthesis or high‑throughput first‑principles calculations, both of which are costly and time‑consuming. Even data‑driven methods that predict properties from known compositions often require a separate optimization loop to propose new candidates, limiting their generative capability. To overcome these bottlenecks, the authors introduce MatGAN, a generative adversarial network (GAN) specifically designed to learn the implicit compositional rules embedded in a large inorganic materials database and to generate novel, chemically plausible inorganic formulas.

Dataset and preprocessing
The authors extracted roughly 120 k inorganic crystal structures from the Inorganic Crystal Structure Database (ICSD). Each entry was reduced to its stoichiometric formula, encoded as a one‑dimensional vector of element fractions. In addition to raw fractions, the model incorporates element‑wise physical descriptors (electronegativity, atomic radius, valence electron count, etc.) via a pre‑trained embedding layer, allowing the network to capture subtle inter‑element relationships. The dataset is split 80 %/10 %/10 % for training, validation, and testing.

Model architecture
MatGAN builds on the Deep Convolutional GAN (DCGAN) framework but adapts it for 1‑D compositional data. The generator receives a 100‑dimensional noise vector sampled from a standard normal distribution and, after several fully‑connected and 1‑D convolutional layers, outputs a vector representing a candidate composition. The discriminator, also a stack of 1‑D convolutions and dense layers, learns to distinguish real ICSD formulas from generated ones. Training uses the Wasserstein GAN with Gradient Penalty (WGAN‑GP) loss to improve stability, and regularization techniques (batch normalization, dropout, learning‑rate scheduling) are applied to avoid over‑fitting.

Training and sampling
After several thousand epochs, the generator is capable of producing compositions that are statistically indistinguishable from the training set in terms of element distribution and stoichiometric patterns. The authors sampled 2 million compositions from the trained generator for evaluation.

Evaluation metrics
Two primary metrics are reported:

1. Novelty – the fraction of generated formulas that do not appear in the original ICSD. MatGAN achieves 92.53 % novelty, indicating that the model is not merely memorizing the training data but is actively exploring new regions of composition space.

2. Chemical validity – the proportion of generated formulas that satisfy charge neutrality and a simple electronegativity‑balance rule (i.e., the difference in electronegativity between cations and anions stays within a reasonable range). Without explicitly encoding these rules, the model produces 84.5 % valid formulas, demonstrating that the discriminator implicitly learns such constraints during adversarial training.

Property prediction and DFT validation
To assess whether the novel, valid candidates could be realistic materials, the authors passed a subset of generated formulas through pre‑trained property‑prediction models (formation energy, band gap, etc.). Several candidates were predicted to have negative formation energies and moderate band gaps, suggesting thermodynamic stability and potential functional utility. For a small selection, density‑functional theory (DFT) calculations were performed; the optimized structures exhibited negative formation energies and no imaginary phonon modes, confirming that at least some GAN‑generated compositions correspond to genuinely stable inorganic compounds.

Discussion of limitations
MatGAN operates solely at the formula level; it does not generate crystal structures or atomic coordinates. Consequently, a generated composition may be chemically valid yet crystallographically infeasible, requiring a downstream structure‑generation step (e.g., using graph‑based generative models or crystal‑structure prediction algorithms). Moreover, the current validity checks are limited to charge neutrality and a rudimentary electronegativity balance; more sophisticated chemical rules such as oxidation‑state consistency, coordination preferences, and bonding topology are not enforced, leaving room for non‑physical outputs. The authors suggest future work incorporating explicit regularization terms for such rules or coupling the GAN with a graph neural network that directly models atomic connectivity.

Implications
By demonstrating that a GAN can learn implicit compositional heuristics from a large inorganic database and generate a massive library of novel, chemically plausible formulas, the study provides a powerful tool for inverse materials design. The high novelty and validity rates imply that MatGAN can dramatically accelerate the candidate‑generation phase of high‑throughput screening pipelines, reducing the computational burden of exhaustive enumeration. When combined with downstream property prediction and structure‑generation modules, this approach could enable rapid discovery of new functional inorganic materials for energy storage, catalysis, electronics, and beyond.

In summary, MatGAN represents a significant step toward data‑driven generative design in inorganic chemistry, showcasing how adversarial learning can capture the hidden “rules of chemistry” embedded in existing databases and leverage them to explore uncharted regions of material space efficiently.