Intelligent Exploration of Thermite Design Space Through Active Learning Algorithms

The reactive materials design space is prohibitively large as not only the chemistry (nature of fuel and oxidizer) but also the microscopic (size of the fuel and oxidizer particles, purity of the metal) and mesoscopic properties (powder density, stoichiometric conditions, binder or additives) influence their macroscopic combustion. For two decades or so, data-driven science is considered as a new paradigm in materials discovery and design that has already achieved number of successes in accompanying and accelerating research and innovation beyond the “Edisonian” approach based on trial-and-error used in most research laboratories. The most used approach is the Bayesian optimization algorithms, which requires large data base for approximating the objective function which is totally not possible when working with reactive materials: each experiment is costly due to the complexity to capture the combustion process <i>in operando</i>, and, truly predictive models that can replace experiments are computationally costly.<br/>In this work, we explore <i>active learning</i> algorithms to allow for a guided and intelligent generation of database starting from a very limited set of available data. As a case study, an Al/CuO powdered thermite (mixture of Al and CuO particles) is considered because it is well-documented from both theoretical and experimental perspectives and a model is available to simulate its combustion ^1,2 in close vessels. A Gaussian process surrogate (GP) is used to approximate the target function in terms of the pressurization rate. Importantly, we explore different acquisition functions translating different exploration tradeoffs from the GP building block to guide the sequential sampling choice (Al/CuO sample) in the database. We study the impact on the quality of the database when prioritizing specific regions of the target space in the acquisition function, while maximizing the coverage of the features space. After presentation of the method and algorithms, we will discuss, (i) the benefit of the proposed algorithm compared to the brute force sampling (systematic sampling) of the Al/CuO design space through LHS technique (Latin Hypercube Sampling) and, (ii) physical aspects of interest of the design space of Al/CuO for gas generation applications.<br/>(1) Tichtchenko, E.; Folliet, V.; Simonin, O.; Bedat, B.; Glavier, L.; Esteve, A.; Rossi, C., Combustion model for thermite materials integrating explicit and coupled treatment of condensed and gas phase kinetics, <i>Proc. Combust. Inst.</i>,<b>2022</b>, 1-9.<br/>(2) Tichtchenko, E.; Bedat, B.; Simonin, O.; Glavier, L.; Gauchard D.; Esteve, A.; Rossi, C., Comprehending the influence of the particle size and stoichiometry on Al/CuO thermite combustion in close bomb: A theoretical study, <i>Propellants, Explosives, Pyrotechnics</i>, <b>2022</b>, e202200334.