# Towards self-driving laboratories in chemistry and materials sciences: The central role of DFT in the era of AI

**Authors:** Bing Huang, Guido Falk von Rudorff, O. Anatole von Lilienfeld  
**arXiv:** [2304.03272](https://arxiv.org/abs/2304.03272)  
**Date:** 2023-04-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.03272v1/

## Key points

- Density functional theory (DFT) is a crucial tool in chemical and materials sciences due to its high predictive power, versatility, and low computational cost.
- Recent progress in machine learning heavily relies on DFT for synthetic data generation and model architecture resulting in models with high efficiency, accuracy, scalability, and transferability (EAST).
- Long-range effects such as charge transfer, conjugation, electron correlation, and London dispersion can propagate throughout the molecule and are modulated by all its local constituents. These effects may not occur in isolation but collectively impact the overall locality of constituents in molecules.
- Short- and long-range effects can become non-separable making it difficult to achieve scalability for large query molecules using atomic representation for molecular description where each atom is represented by a vector encoding many-body interactions between it and its neighbors.
- Accurately accounting for long-range interactions during ML training is difficult since they are much smaller than the target size.
- Despite these challenges, recent developments have paved the way for routine use of successful experimental planning software within self-driving laboratories.
- The Materials Project is an example of how AI can accelerate materials innovation by providing open access to a vast database containing information about materials properties obtained from DFT calculations.
- Other initiatives like QM7-x, Improved Decision Making with Similarity Based Machine Learning, and Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules demonstrate how AI can be used to design new molecules with desired properties.
- In conclusion, DFT plays a central role in the era of AI for chemistry and materials science.

## Summary

Density functional theory (DFT) has become a crucial tool in the chemical and materials sciences due to its high predictive power, versatility, and low computational cost. Recent progress in machine learning (ML) model development heavily relies on DFT for synthetic data generation and model architecture, resulting in models with high efficiency, accuracy, scalability, and transferability (EAST). However, the many-body interactions between atoms in molecules pose challenges for scalable ML models. Long-range effects such as charge transfer, conjugation, electron correlation, and London dispersion can propagate throughout the molecule and are modulated by all its local constituents. These effects may not occur in isolation but collectively impact the overall locality of constituents in molecules. As a result, short- and long-range effects can become non-separable. To achieve scalability for large query molecules using atomic representation for molecular description where each atom is represented by a vector encoding many-body interactions between it and its neighbors is necessary but not sufficient. The lack of scalability remains one of the most common issues of state-of-the-art ML models. Accurately accounting for long-range interactions during ML training is difficult since they are much smaller than the target size. Among long-range effects of quantum mechanical origin such as conjugation effects and electron correlation, decent scalable ML-based approximations have yet to be invented. Despite these challenges, recent developments have paved the way for routine use of successful experimental planning software within self-driving laboratories. This will enable chemists to focus more on designing experiments rather than performing them manually. The Materials Project is an example of how AI can accelerate materials innovation by providing open access to a vast database containing information about materials properties obtained from DFT calculations [124]. Other initiatives like QM7-x [125], Improved Decision Making with Similarity Based Machine Learning [126], and Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules [127] demonstrate how AI can be used to design new molecules with desired properties. In conclusion, DFT plays a central role in the era of AI for chemistry and materials science.