Machine Learning for Munitions Development

It takes 10+ years for next generation steels to progress through discovery to engineering application.  The reasons for long development timelines are many and include risk adverse engineering practices, limited understanding of the fundamental knowledge base, length scales, screening experiments, diagnostic techniques and customer-based needs.  This proposal will examine how the timeline for the development of multi-phase steel systems can be contracted by the exploitation of machine learning (ML) and artificial intelligence (AI).  The overarching goal is to capture elements of the structure-processing-properties paradigm of material science and utilize the data, data descriptors and the algorithms of ML/AI in the most responsive way to move digitally through element selections, thermo-mechanical treatment and into property evaluation.  Figure 1 graphically captures the notion of relationships within this structure-processing-properties paradigm.  For DoD munitions developers the success of a penetrator case steel for a blast/fragmentation munitions requires a trade space analysis of high strength, high toughness, impact resistance, shock, fragmentation and affordability.  The key to establishing a robust analysis tool is the determination of the critical microstructural linkages to these performance goals.  With microstructural linkages providing descriptors, algorithms from the field of optimization, or AI will be utilized to digitally probe the design space.  Several techniques exist for analysis of data sets that include regression, Bayesian, Neural Network (NN), Deep Neural Network (DNN), Genetic Algorithms (GA), etc.  The learning process in digital engineering is pursued by creating perturbations to the initial set of variables and then observing their impact on the objective, acquisition or fitness test relative to the desired goals of minimizing or maximizing these functions.  New data from these steps in the calculation will be fed back into the process as updates to the original information and used to further improve subsequent solutions.  These processes normal iterate to a desire level of success in the objective function, or until a maximum number of iterations have been completed.  In the Phase I portion of this investigation, IS4S proposes to review the available material characterization data sets and their architecture from the perspective of use in ML/AI.  Specific test case will be made of 4340, a high strength low alloy martensitic steel.  A material that has found numerous applications in munitions because of the balance of strength, toughness and affordability.