How to do impactful research in artificial intelligence for chemistry and materials science

Austin H. Cheng, Cher Tian Ser, Marta Skreta, Andrés Guzmán-Cordero, Luca A. Thiede, Andreas Burger, Abdulrahman Aldossary, Shi Xuan Leong, Sergio Pablo-García, Felix Strieth-Kalthoff, Alán Aspuru-Guzik

Abstract

Machine learning has been pervasively touching many fields of science. Chemistry and materials science are no exception. While machine learning has been making a great impact, it is still not reaching its full potential or maturity. In this perspective, we first outline current applications across a diversity of problems in chemistry. Then, we discuss how machine learning researchers view and approach problems in the field. Finally, we provide our considerations for maximizing impact when researching machine learning for chemistry.

Introduction

While machine learning (ML) has become pervasive in chemistry and materials science, a significant gap remains between the ML community—which values algorithmic novelty and benchmark performance—and the chemistry community, which values physical insight and experimental validation. Many researchers struggle to identify which ML tools are appropriate for specific chemical problems, often leading to "fantasy" molecular designs that cannot be synthesized or models that fail to generalize beyond narrow datasets. This paper seeks to bridge this gap by providing a shared language and a framework for choosing projects that solve essential, long-standing scientific questions.

A Taxonomy of Chemical AI Problems

To organize the field, the authors propose a conceptual framework that categorizes chemical AI research into five main flows centered around molecular structure and its relationship to other domains.

A conceptual map of machine learning applications in chemistry, illustrating the flows between structure, property, physics, and analysis.

Key components of this taxonomy include:

  • Structure-to-Property (Prediction): Mapping chemical structures to observable properties.
  • Property-to-Structure (Inverse Design): Generating structures that satisfy specific property requirements.
  • Structure-to-Structure (Synthesis Planning): Predicting reaction pathways and outcomes.
  • Structure-to-Physics (Simulation): Using ML to accelerate physical simulations and 3D structure generation.
  • Structure-to-Analysis (Spectroscopy): Bridging the gap between experimental data (like spectra) and chemical identity.

These tasks are further mapped onto standard ML problem classes such as regression, generative modeling, sampling, and agent-based optimization.

Strategic Project Selection

To maximize impact, the authors adapt the "Aspuru-Guzik/Whitesides rules," posing three critical questions for any new project:

  1. Has this been solved before? Researchers should aim to be first to apply a new ML framing to an underexplored field (e.g., agrochemistry) rather than achieving marginal gains on saturated benchmarks.
  2. Is it relevant to society? Focus on problems with real-world implications, such as sustainability or drug discovery.
  3. Is it feasible? Ensure the project is achievable with available computational and experimental resources.

Datasets and the Data Gap

The paper emphasizes the importance of benchmarks like GuacaMol and MOSES for molecular design, and the Open Catalyst Project (OCP) for materials science. However, a significant challenge remains: unlike the internet-scale data available for language models, high-quality chemistry data is scarce and expensive. This necessitates the rise of Self-Driving Labs (SDLs) to generate autonomous experimental data and the use of the Open Reaction Database (ORD) to digitize chemical knowledge.

Key Insights and Ideas

  • The "Bitter Lesson" in Chemistry: While domain heuristics provide short-term gains, scale and compute typically win in the long run. However, chemistry requires a unique balance because data is so hard to acquire.
  • Reframing Search as Generation: Molecular discovery should be treated as a generative sampling problem to navigate the astronomical chemical space ($10^{33}$ to $10^{60}$ molecules) rather than a traditional "needle in a haystack" search.
  • Foundation Models Need Physics: Impactful foundation models must be multi-modal, incorporating 3D geometry and physical constraints rather than relying solely on text-based literature.

Impact and Future Directions

ML has already enabled massive scale in discovery, such as screening $10^7$ candidates in the Harvard Clean Energy Project and doubling reaction yields through ML-driven optimization. Future work must address the synthesizability barrier, where generative models propose molecules that are chemically valid but impossible to make. Integrating synthesis planning and robotics into the design loop is essential. Furthermore, the field must move beyond small organic molecules toward complex mixtures and heterogeneous materials. The ultimate goal is the development of "universal" models capable of handling multiple modalities—structure, text, and spectra—simultaneously.

Summary updated

Abstract

Machine learning has been pervasively touching many fields of science. Chemistry and materials science are no exception. While machine learning has been making a great impact, it is still not reaching its full potential or maturity. In this perspective, we first outline current applications across a diversity of problems in chemistry. Then, we discuss how machine learning researchers view and approach problems in the field. Finally, we provide our considerations for maximizing impact when researching machine learning for chemistry.
BibTeX
@inproceedings{cheng2024how,
  author = {Austin H. Cheng and Cher Tian Ser and Marta Skreta and Andrés Guzmán-Cordero and Luca A. Thiede and Andreas Burger and Abdulrahman Aldossary and Shi Xuan Leong and Sergio Pablo-García and Felix Strieth-Kalthoff and Alán Aspuru-Guzik},
  title = {{How to do impactful research in artificial intelligence for chemistry and materials science}},
  booktitle = {Faraday},
  year = {2024},
  doi = {10.1039/d4fd00153b},
}
Figure