November 7, 2023
By Anatole von Lilienfeld
The introduction of AI into the mix of physics-based computer simulations has led to fascinating changes in the world of chemistry and materials science. Instead of spending ages poking around in the dark studying one compound at a time, we’ve now got machine learning gadgets sorting through mountains of virtual compounds, providing reliable answers about new chemical compounds, or new ways to tweak existing materials for special jobs.
This change is akin to the introduction of calculators replacing longhand calculations. However, in this case, the calculators give us a tremendous shortcut to understanding and control. It could soon become routine to craft the very stuff our world is made of on-demand and in real-time, allowing us to tackle many of the world’s challenges through improved materials and molecules. Imagine how this will allow us to accelerate drug discovery and personalized medicine, or reduce our carbon footprint by creating novel battery technologies.
I find it remarkably beautiful to blend these old and new ways of thinking, an approach exemplified by the following three publications from my lab.
Quantum mechanics, especially an approach called density functional theory (DFT), is a powerful tool in understanding chemicals and materials and predicting their properties and behavior. This method is prized for its accuracy, versatility, and universality in computer simulations and is broadly applicable across Mendeleev’s periodic table. We have recently made a lot of advancements in using machine learning (ML) models trained on synthetic data that was obtained from computationally demanding DFT simulations. Already in 2011, we could show that after training, these ML models could deliver instant predictions for novel compounds with DFT quality and versatility. More recent advancements such as generative AI and large language models have led to further major impact and are setting the stage for software that holds great promise for planning successful experimental validations in labs that could essentially run themselves.
In the paper “The central role of density functional theory in the AI age” published in Science, we discuss how this enables a future where robot-led experiments become as fundamental to science as machine learning, computer simulations, traditional theories, and human-led experiments.
Within the realm of our quantum mechanics-based ML models, the way we represent or “map” chemical systems is crucial since it directly affects the training data needs required to reach desirable predictive power. The most predictive methods of representation, which get away with minimal training data set sizes, however, tend to be computationally heavy, making the training step slow and taxing in terms of hardware needs and carbon footprint.
Since moving my lab to Toronto last summer, Vector postdoctoral fellow Stefan Heinen and Vector grad student Danish Khan and I have been working on a novel physics-based featurization of chemical compounds that is so compact that the training cost of ML models can be reduced by multiple orders of magnitudes. In our paper “Kernel based quantum machine learning at record rate: Many-body distribution functionals as compact representations,” which was published in The Journal of Chemical Physics, we propose an improved way to represent chemical systems that’s not only ultra-compact but that is also invariant with respect to the system’s size. It’s based on something dubbed atomic Gaussian many-body distribution functionals (MBDF).
When we tested MBDF on benchmark data sets consisting of organic molecules, its performance matched or even rivaled the best current methods for various quantum properties. Our findings suggest that the MBDF-based approach can efficiently navigate the balance between the cost of sampling and training while maintaining high accuracy, making it the preferred choice for certain settings of available training data and compute hardware. Generating chemically accurate predictions for quantum properties of unseen out-of-sample compounds, it achieves sampling rates in the chemical compound space that amount to sifting through roughly 48 molecules/second using just a single compute core. While humans are typically incapable of providing quantitative estimates of quantum properties, numerically solving the corresponding equations for just a single molecule would easily consume thousands of seconds when done the conventional way without ML.
This summer Siwoo Lee, an undergraduate in the Department of Chemistry at University of Toronto, Heinen, Khan, and I developed an autonomous workflow that combines a convolutional neural network with a large language model to pull specific tables of data from scientific papers. We tested this approach for 592 organic molecules that were studied within 74 different papers published between 1957 and 2014. These papers reported experimental measurements on a property crucial for electrochemistry research, the oxidation potential, with values ranging from -0.75 to 3.58 V and the sign indicating if the molecule would rather attract or release an electron.
After curating the data for validation and to account for differing experimental conditions, we used it to train additional machine learning models which were able to predict oxidation potentials with an error margin close to what you’d expect from regular experimental errors (about +/- 0.2 V). If multiple studies had results for the same molecule, our AI model could decide which value was most likely the correct one. We then used our models to predict the oxidation potential for over 100,000 organic molecules, finding values between 0.21 to 3.46 V. Our analysis showed that certain molecule features, like being aliphatic, could raise the oxidation potential from an average of 1.5 to 2.0 V, while having more atoms generally lowered it. Importantly, our workflow demonstrates how daisy chaining multiple AI models enables an automatic workflow that will significantly cut down the manual work scientists would normally have to do to obtain computational property estimates of novel compounds based on AI models trained on literature data.
This exciting new work has been submitted for publication and is currently available online as a pre-print. Future applications of this line of research might well contribute to the revolution of experimental materials and chemistry research efforts through autonomous AI agents investigating novel questions and problems through the use of self-driving laboratories.
We’re on the brink of some really neat stuff; we could soon be stumbling upon new medicines, crafting better batteries, producing improved organic electronics, concocting cleaner ways to drive chemical reactions with tailored catalysts, and maybe, just maybe, chancing on those elusive room temperature superconductors.
