In collaboration with researchers at APL, Drs. Nam Q. Le and Jarod Gagnon, the Clancy group will use an “on-the-fly” active learning approach within the umbrella of machine learning to study a novel additive manufacturing process to create gallium nitride thin films. This approach combines the accuracy of a first-principles, ab initio, method with the orders of magnitude faster execution speed of using an empirical force field MD, essentially the best of both worlds. It will also us to capture the details of the formation of gallium nitride by a chemical reaction in the liquid phase and model the subsequent crystallization process with atomistic precision. Being able to essentially ‘print’ this material should have implications for energy transmission and efficiency.
We used an “on-the-fly” active learning approach, a machine learning technique, to simulate a novel additive manufacturing (AM) process to create high-value gallium nitride thin films. Being able to essentially ‘print’ this material should have implications for improved energy transmission and efficiency. Such an approach can provide a level of atomic-scale detail on the crystallization process, e.g., identifying the fastest growing facets, that is unavailable by experimental means. This kind of information would be invaluable for providing a level of control of the AM process that doesn’t currently exist.
In principle, Molecular Dynamics (MD) is a suitable way to obtain this kind of atomic scale information. But its accuracy is limited either by the inherent inadequacies of using an empirical force field to solve the central F=ma equation of motion in MD, especially for a system involving a chemical reaction like the AM process. Or it is limited by the extreme computational cost of using an ab initio approach to obtain those forces more accurately. The practical value of an “on-the-fly” approach is that it combines the accuracy of an ab initio, quantum mechanical, method within an MD framework. This is essentially the best of both worlds: Active learning allows the system to “learn” where next to move the particles without invoking the high cost of performing an ab initio calculation unless the algorithm detects that it is deviates sufficiently from its training set that it needs to make an ab initio calculation to stay accurately on track. In this way, we can capture the details of the formation of gallium nitride by a chemical reaction of gaseous nitrogen atoms diffusing through liquid gallium to form crystalline GaN.
In this seed award, we are focusing on following the ability of the active learning algorithm to capture this chemical reaction process and ensure that the solid that it produces is both crystalline and has an expected crystal structure. The algorithm we are using is known as FLARE, Fast Learning of Atomistic Rare Events, developed by Kozinsky (Harvard) in 2020. It is a Bayesian approach based on a Gaussian process regression technique. FLARE is not a “turn-key” system and it has yet to be tested on a system that involves a reaction-diffusion system and three phases of matter (solid, liquid, gas). So far in this seed award, we have used well-known Steinhardt Order Parameters that provide a unique “fingerprint” for known crystal structures (e.g., FCC, BCC, etc.) based on the angles between nearest neighbors of a given atom. This allows you to see the evolving crystal structure as the liquid crystallizes, which may not follow a simple monolithic path. Figure 1 (a-c) shows how the order parameters, q4 and q6, evolve from values representative of a random liquid (shown in 1a at time t=0) towards that of pure HCP (hexagonal close packed structure) shown as a black dot at the top RH corner of the plot. The crosses show order parameters from each layer. These get closer to pure HCP as time progresses from 1 ns I(Fig. 1b) to 2 ns (Fig.1c). The simulations are continuing for longer times to bring the structure to completion as an HCP crystal characteristic of GaN.
PI: Paulette Clancy, PhD (Department Head, Professor, Department of Chemical and Biomolecular Engineering)
Co-Is: Nam Q. Le and Jarod Gagnon
Paulette Clancy is a Professor and inaugural Head of the Department of Chemical and Biomolecular Engineering at Johns Hopkins University. Her research group is recognized as one of the country’s leading computational groups in atomic-scale modeling of materials and algorithm development. Her current thrust is to develop machine learning algorithms to accelerate the search for optimal materials processing protocols.
