Scientists Fuse Simulations and Machine Learning to Accelerate Novel Additively Manufactured Materials

Researchers at the Johns Hopkins Applied Physics Laboratory have developed a machine learning method that could have a huge impact on understanding how material is formed during the additive manufacturing process.

December 1, 2024

John Hopkins Applied Physics Laboratory, Laurel, MD

APL researchers developed a precise modeling approach to simulate critical thermal dynamics. The melt pool microstructure of a precipitation-strengthened nickel-base superalloy, shown here, was used to validate APL’s model and cooling rate prediction. (Image: ASM International, 2024.)

Researchers at the Johns Hopkins Applied Physics Laboratory (APL) in Laurel, Maryland, have demonstrated a novel approach for applying machine learning to predict microstructures produced by a widely used additive manufacturing technique. Their approach promises to dramatically reduce the time and cost of developing materials with tailored physical properties and will soon be implemented on a NASA-funded effort focused on creation of a digital twin.

“We anticipate that this new approach will be extremely impactful in helping design and understand material formation during additive manufacturing processes, and this fits into our overarching strategy focused on accelerating materials development for national security,” said Morgan Trexler, who manages APL’s Science of Extreme and Multifunctional Materials program in the Research and Exploratory Development Mission Area.

The modeling approach focuses on laser powder bed fusion (LPBF), in which layers of metal powder are fused to create 3D objects with a high-powered laser. The process can produce strong, dense metal parts in complex geometries. However, there are many possible variations in processing conditions, resulting from the metal powder characteristics, laser settings and the interactions between these. As a result, the properties of the final materials can vary widely. The key to making LPBF successful, then, is the ability to predict the microstructure of the printed component before printing it, which in turn can be used to predict the mechanical properties of the material and the performance of the fabricated component.

Led by Li Ma, an APL Senior Staff Engineer whose specialties include integrated computational materials modeling for additive manufacturing, researchers developed a technique to predict the microstructure formed during a single laser pass over a certain volume of powder via physics-based computational modeling and simulation.

Specifically, the team used a computational fluid dynamics (CFD) model to quantify changes in temperature and cooling rates during the printing process. Thermal gradient and cooling rate from CFD were used as input into another model that predicts microstructural formation, including grain structure and phase formation. In the future, it will be possible to use these simulations to predict the microstructures of fully manufactured components.

It was a monumental achievement — but it didn’t go quite far enough. Manufacturing a single component entails millions of laser passes. Simulating each small region of that component’s microstructure would take hours of high-performance computing time, severely diminishing the practical advantages of using simulations in the first place.

So, team member Hudson Liu, an APL Student Program to Inspire, Relate, and Enrich (ASPIRE) intern and high school student at Gilman School in Baltimore, developed the final puzzle piece: a machine learning model that would greatly reduce the need to run costly simulations.

Liu combined several existing machine learning model types to create what the team refers to as a diffusion probabilistic field model. Somewhat like the familiar Stable Diffusion platform that creates images from text, Liu’s model generates images based on two key LPBF parameters: the cooling rate and the temperature gradient, which quantifies how temperature changes over distance, particularly from the point of laser impact and the surrounding solid metal.

The model was trained on data from more than 400 simulations conducted at APL, with the microstructure model validated against experimental microscopy measurements of the LPBF material provided by Maureen Williams and Lyle Levine at the National Institute of Standards and Technology in Gaithersburg, Maryland.

The result is a model that can effectively approximate simulations with enough accuracy to dramatically accelerate the process of developing additively manufactured materials using LPBF.

“The key benefit of using a model is its speed. Our model can approximate in seconds or minutes what would take hours in a simulation,” Liu said. “This allows researchers to quickly explore a wide range of parameters and at much lower cost.”

Originally developed with internal funding, this work is directly applicable to NASA’s Space Technology Research Institute (STRI), and the team is working to implement it soon toward that project’s goals.

“NASA wants validated models that can help them predict what will happen in a build, and how the subsequent part will perform, without expensive experimentation,” Ma said. “So this approach is valuable, particularly when you think about doing additive manufacturing on the Moon or in space, where experimentation becomes so expensive that it’s effectively impossible.”

This work serves as a link in a chain in which computational scientists on the STRI team are working to predict the properties of full-size additively manufactured aerospace components directly from the characteristics and processing conditions of the raw materials.

The team is already at work scaling up what their models can do. In addition to predicting microstructures of larger components and the results of more laser passes, they intend to train models using video data, which will enable them to predict 3D as well as 2D microstructures.

The team’s work was published in a special issue of Metallography, Microstructure, and Analysis. Ma presented results from the paper at the Integrated Computational Materials and Engineering for Defense Summit and at APL’s Accelerating Materials Discovery for National Security Symposium in September. She is also chairing sessions at the Materials Science and Technology technical meeting and exhibition in October.

This effort is part of a larger portfolio of work at APL focused on using artificial intelligence to accelerate the discovery of novel materials for extreme environments.

This article was written by Ajai Raj, Science and Technology Writer, The Johns Hopkins University Applied Physics Laboratory. For more information, contact Amanda Zreibec, This email address is being protected from spambots. You need JavaScript enabled to view it..