From ChatGPT to DALL-E, the realm of deep learning artificial intelligence (AI) is expanding into diverse fields. A recent investigation by researchers from the University of Toronto Engineering, detailed in Nature Communications, challenges a core assumption of deep learning models – the belief that they necessitate vast amounts of training data.
Professor Jason Hattrick-Simpers and his team are immersed in advancing next-gen materials, ranging from catalysts transforming captured carbon into fuels to creating non-stick surfaces for ice-free airplane wings.
A significant challenge lies in navigating the vast potential search space. The Open Catalyst Project, for instance, boasts over 200 million data points on potential catalyst materials. This covers only a fraction of the immense chemical space, potentially concealing catalysts crucial for addressing climate change.
Hattrick-Simpers states, “AI models can efficiently navigate this space, narrowing down choices to the most promising material families.” He underscores the need to identify smaller datasets for equitable access, avoiding the requirement for supercomputers.
Yet, a second challenge emerges. Existing smaller materials datasets are often tailored to specific domains, potentially limiting diversity and missing unconventional yet promising options.
Dr. Kangming Li, a postdoctoral fellow in Hattrick-Simpers’ lab, likens this to predicting students’ grades based on previous test scores from a specific region. The challenge in materials research mirrors this, requiring consideration of global diversity.
One potential solution involves identifying subsets within large datasets that are easier to process while retaining crucial information and diversity. Li developed methods to identify high-quality subsets from databases like JARVIS, The Materials Project, and the Open Quantum Materials Database.
Li’s computer model, trained on the original dataset and a 95% smaller subset, yielded intriguing results. Predicting properties within the dataset’s domain showed comparable performance, suggesting that more data doesn’t necessarily enhance model accuracy. This highlights potential redundancy in large datasets.
The findings underscore that even models trained on smaller datasets can excel with high-quality data. Hattrick-Simpers emphasizes the nascent stage of using AI for materials discovery, urging careful consideration in dataset construction.
The key takeaway is the necessity for thoughtful dataset construction, focusing on information richness rather than sheer volume, a critical aspect as AI continues to revolutionize materials science.