The team, led by scientists from the University of Cambridge, developed a 'multi-fidelity' machine-learning approach to predict the properties of materials. This method combined information about the material’s structure obtained using computational techniques with experimental measurements to build accurate predictive models using deep-learning algorithms.
The scientists tested the multi-fidelity approach on four materials: steel alloys, high-entropy alloys, thermoelectric materials, and metal-organic frameworks. They demonstrated that their method achieved state-of-the-art performance in predicting the properties of these materials.
For example, for steel alloys, the multi-fidelity model predicted the material's yield strength with a mean absolute error (MAE) of only 1.8%, compared to 4.5% for the best previous method. For high-entropy alloys, the multi-fidelity model predicted the material's Vickers hardness with an MAE of 2.3%, compared to 5.8% for the best previous method.
"Machine-learning techniques can predict the properties of materials and significantly reduce the time and cost of materials discovery,” said co-first author Dr. Hao Wu from the Department of Materials Science & Metallurgy at Cambridge. “But for machine learning to provide high-fidelity, computationally efficient predictions, we must combine multiple complementary information sources, such as physics-informed models and experimental measurements.”
Materials discovery and development currently involve an iterative cycle of material synthesis, experiments to measure material properties, and costly computational simulations to understand the underlying mechanisms. This approach is time consuming, expensive, and inefficient, and it typically requires human experts with deep knowledge of physics or chemistry.
The new multi-fidelity machine-learning approach streamlines the design process by efficiently identifying the most promising material candidates without having to perform many time-consuming experiments or high-fidelity computations.
"A typical high-fidelity computational simulation may take a week or even months to complete,” said co-first author Dr. Xiaoqing Huang from the Department of Materials Science & Metallurgy. “If we want to explore hundreds of materials, it’s practically infeasible to obtain high-fidelity computational results for all of them. Our multi-fidelity deep-learning framework overcomes this by utilizing low-cost physics-based simulations and experimental measurements to guide the learning of high-fidelity models.”
By significantly reducing the time and cost associated with materials discovery, the new multi-fidelity machine-learning technique can accelerate the development of new and improved materials for a wide range of applications, including energy storage, catalysis, and aerospace.
"We believe that our approach can not only enable breakthroughs in materials discovery and development but also benefit computational science and design in other disciplines, such as chemistry, biology and pharmaceutical research,” said senior author Professor Li Yang from the Department of Materials Science & Metallurgy. “We hope that this work will pave the way for integrating multi-scale simulations and experiments in a data-driven framework for scientific research and engineering design.”
