[논문 리뷰] Generative artificial intelligence for computational chemistry: a roadmap to predicting emergent phenomena

Pratyush Tiwary, Lukas Herron외 4명|arXiv (Cornell University)|2024. 09. 04.

Scientific Computing and Data Management인용 수 5

한 줄 요약

This Perspective surveys Generative AI methods for computational chemistry, outlining theoretical foundations, current methods, applications, and the challenges to predicting emergent phenomena.

ABSTRACT

The recent surge in Generative Artificial Intelligence (AI) has introduced exciting possibilities for computational chemistry. Generative AI methods have made significant progress in sampling molecular structures across chemical species, developing force fields, and speeding up simulations. This Perspective offers a structured overview, beginning with the fundamental theoretical concepts in both Generative AI and computational chemistry. It then covers widely used Generative AI methods, including autoencoders, generative adversarial networks, reinforcement learning, flow models and language models, and highlights their selected applications in diverse areas including force field development, and protein/RNA structure prediction. A key focus is on the challenges these methods face before they become truly predictive, particularly in predicting emergent chemical phenomena. We believe that the ultimate goal of a simulation method or theory is to predict phenomena not seen before, and that Generative AI should be subject to these same standards before it is deemed useful for chemistry. We suggest that to overcome these challenges, future AI models need to integrate core chemical principles, especially from statistical mechanics.

연구 동기 및 목표

• Summarize the theoretical concepts bridging Generative AI and computational chemistry.
• Review widely used Generative AI methods and their chemistry-relevant variants.
• Highlight selected applications in ab initio chemistry, force fields, and biomolecular structure prediction.
• Discuss the challenges to predictive utility, especially emergent phenomena, and propose a research roadmap.

제안 방법

• Describe autoencoders and derived methods with physics-informed losses and latent priors.
• Explain GANs and variants (cGANs, WGANs) and their chemistry limitations.
• Summarize reinforcement learning approaches and issues (curse of dimensionality, data scarcity, mode collapse).
• Present flow-based methods and diffusion models, including normalizing flows and score-based diffusion.
• Cover recurrent architectures and large language models, including AlphaFold-related use in structure prediction and limitations.

실험 결과

연구 질문

• RQ1 How can Generative AI methods be integrated with core chemical principles and statistical mechanics to predict emergent phenomena?
• RQ2 What are the strengths, limitations, and trade-offs of prominent Generative AI methods (AE/VAE, GANs, RL, flows, LLMs) in molecular modeling?
• RQ3 In which chemistry applications (QM, force fields, protein/RNA structure) do these methods show the most promise, and why?
• RQ4 What strategies are needed to overcome data, generalization, and physicality challenges when applying Generative AI to chemistry?
• RQ5 What roadmap steps are required before Generative AI becomes a reliable tool for predicting emergent chemical phenomena?

주요 결과

• Generative AI methods have potential to sample structures, develop force fields, and accelerate simulations in chemistry.
• Current AI tools often struggle with emergent phenomena and tend to excel at memorization or interpolation rather than true prediction.
• GANs face training instability and mode collapse, leading to slower adoption in chemistry relative to diffusion models and RL-based approaches.
• Flow-based and diffusion models offer promising, Jacobian-aware or score-based sampling strategies that align with statistical physics ideas.
• LLMs and diffusion-based RNN/transformer approaches extend structure and ensemble prediction but face extrapolation and physicality challenges when applied to chemistry.
• A central conclusion is that integrating core chemical principles, especially from statistical mechanics, is essential for reliable chemical predictions.