Abstract:Generative artificial intelligence offers a promising avenue for materials discovery, yet its advantages over traditional methods remain unclear. In this work, we introduce and benchmark two baseline approaches - random enumeration of charge-balanced prototypes and data-driven ion exchange of known compounds - against three generative models: a variational autoencoder, a large language model, and a diffusion model. Our results show that established methods such as ion exchange perform comparably well in generating stable materials, although many of these materials tend to closely resemble known compounds. In contrast, generative models excel at proposing novel structural frameworks and, when sufficient training data is available, can more effectively target properties such as electronic band gap and bulk modulus while maintaining a high stability rate. To enhance the performance of both the baseline and generative approaches, we implement a post-generation screening step in which all proposed structures are passed through stability and property filters from pre-trained machine learning models including universal interatomic potentials. This low-cost filtering step leads to substantial improvement in the success rates of all methods, remains computationally efficient, and ultimately provides a practical pathway toward more effective generative strategies for materials discovery.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to evaluate the performance of generative artificial intelligence (AI) in the discovery of inorganic crystalline materials and compare it with traditional methods in order to clarify the advantages and limitations of generative models. Specifically, the author explores the following issues by introducing two baseline methods - randomly enumerating charge - balanced prototype structures and data - driven ion - exchange of known compounds - and benchmarking them against three generative models (variational auto - encoder, large - language model, and diffusion model): 1. **Are generative models more effective in generating stable materials than traditional methods?** Evaluate the performance of generative models in this regard by comparing the thermodynamic stabilities of materials generated by different methods. 2. **Can generative models propose more novel structural frameworks?** Examine the novelty of generative models in generating new materials, especially their ability to explore unstudied chemical compositions and structures. 3. **Can generative models effectively optimize for specific properties while maintaining a high stability rate?** For example, physical properties such as electronic bandgap and bulk modulus. To enhance the performance of these methods, the author also introduced a post - generation screening step, that is, all proposed structures are filtered for stability and properties by pre - trained machine - learning models (including general inter - atomic potentials). This step not only improves the success rate of all methods but also maintains computational efficiency, providing a practical path for more effective generation strategies. ### Main Conclusions - **Traditional methods such as ion - exchange perform well in generating stable materials**, but many of the generated materials are similar to known compounds. - **Generative models have an advantage in proposing novel structural frameworks**, especially when there is sufficient training data, they can more effectively optimize for specific properties while maintaining high stability. - **Combining the post - generation screening step can significantly improve the success rate of all methods**, and it is still computationally efficient. Through these evaluations, this study provides important baselines and reference standards for the future development and application of generative models.

Establishing baselines for generative discovery of inorganic crystals

Unleashing the power of novel conditional generative approaches for new materials discovery

3-D Inorganic Crystal Structure Generation and Property Prediction via Representation Learning

Generative Hierarchical Materials Search

Generative Models for Automatic Chemical Design

MatterGen: a generative model for inorganic materials design

Exploration of crystal chemical space using text-guided generative artificial intelligence

Scalable Diffusion for Materials Generation

Self-Supervised Generative Models for Crystal Structures

High‐Throughput Discovery of Novel Cubic Crystal Materials Using Deep Generative Neural Networks

Data-driven discovery of 2D materials by deep generative models

Deep Generative Models for Materials Discovery and Machine Learning-Accelerated Innovation

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

Evaluating the diversity and utility of materials proposed by generative models

Data-driven discovery of novel 2D materials by deep generative models

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

Generative Design of inorganic compounds using deep diffusion language models

Generative Design of Crystal Structures by Point Cloud Representations and Diffusion Model

Data-Driven Score-Based Models for Generating Stable Structures with Adaptive Crystal Cells

Generative Inverse Design of Crystal Structures via Diffusion Models with Transformers

Dismai-Bench: Benchmarking and designing generative models using disordered materials and interfaces