Estimating the thermal insulating performance of multi-component refractory ceramic systems based on a machine learning surrogate model framework

Abstract

Predicting the insulating thermal behavior of a multi-component refractory ceramic system could be a difficult task, which can be tackled using the finite element (FE) method to solve the partial differential equations of the heat transfer problem, thus calculating the temperature profiles throughout the system in any given period. Nevertheless, using FE can still be very time-consuming when analyzing the thermal performance of insulating systems in some scenarios. This paper proposes a framework based on a machine learning surrogate model to significantly reduce the required computation time for estimating the thermal performance of several multi-component insulating systems. Based on an electric resistance furnace case study, the framework estimated the feasibility and the final temperature of nearly 1.9×105 insulating candidates’ arrangements with reasonable accuracy by simulating only an initial sample of 2.8% of them via FE. The framework accuracy was evaluated by varying the initial sample size from ≈0.9% to 8% of total combinations, indicating that 3%–5% is the optimal range in the case study. Finally, the proposed framework was compared to the evolutionary screening procedure, a previously proposed method for selecting insulating materials for furnace linings, from which it was concluded that the machine learning framework provides better control over the number of required FE simulations, provides faster optimization of its hyperparameters, and enables the designers to estimate the thermal performance of the entire search space with small errors on temperature prediction.

Theory-guided design of high-strength, high-melting point, ductile, low-density, single-phase BCC high entropy alloys

“…The cost of exploring the entire phase space with only historical data and intuitive knowledge is inaccessibly high. Database and machine learning methods [3][4][5][6][7][8] hold promise but the amount of data available is paltry compared to the typical scales on which machine learning is applied in other domains. Finding new high-performance alloys must thus rely on a physics − based understanding of the connection between properties and composition.…”