Use of machine learning and predictive analytics for modelling and simulation of the degradation of Ultra High Temperature Ceramics (UHTCs) under hypersonic flow

This thesis supports research on the use of machine learning (ML) and predictive analytics to model and simulate the degradation of Ultra-High Temperature Ceramics (UHTCs) under hypersonic flow conditions. UHTCs are crucial for protecting leading edges and key parts of hypersonic vehicles due to their capacity to resist extreme temperatures, typical of hypersonic flight. The study aims to improve the accuracy and efficiency of degradation modelling to aid design and enhance the performance and reliability of these components.

Zirconium and hafnium-based samples refer to ultra-high temperature ceramic (UHTC) composites that incorporate zirconium (Zr) and hafnium (Hf) compounds, primarily in the form of diborides (ZrB2, HfB2) with silicon carbide (SiC) additives. These materials are extensively studied for their high thermal stability, oxidation resistance, and mechanical strength under extreme conditions, such as hypersonic flight and space re-entry. ZrB2–SiC composites exhibit high thermal conductivity and generally good oxidation resistance due to the formation of a protective SiO₂-rich glass layer. However, at elevated temperatures and under harsh oxidising or water-vapour-rich environments, this silica-based layer can volatilise, and the associated ZrO₂ scale may crack, reducing long-term protection. In contrast, HfB2–SiC composites demonstrate superior oxidation resistance and higher temperature capability, as the resulting HfO₂ layer is more stable and adherent. Nevertheless, this improvement comes with increased density, which is a drawback for weight-sensitive aerospace applications.

To analyse these degradation mechanisms, a bespoke mathematical model was developed to estimate porosity evolution and oxidation-induced strain, revealing deformation regions of 0.2mm and height variations ranging from 5 to 25mm under extreme aerothermal conditions. Additionally, stress distribution analysis using ML techniques such as Principal Component Analysis (PCA), K-means clustering, and topological mapping identified localised stress concentrations of approximately 400 MPa in critical areas of ZrB2-SiC composites. The integration of ML-driven predictive analytics significantly improved the accuracy of stress distribution and porosity estimation, demonstrating computational efficiency gains over traditional Finite Element Modelling (FEM) methods. This study highlights the potential of ML in expediting material discovery and enhancing simulation accuracy, ultimately contributing to the development of reliable, non-ablative thermal protection systems for aerospace applications. Future work includes validating ML predictions with empirical data, exploring advanced ML algorithms, and extending the framework to other aerospace materials.

To accelerate Computational Fluid Dynamics (CFD), a surrogate framework was introduced. A total of 10,000 Monte Carlo simulations combined with two-temperature modelling were executed using Pyfoam to generate training datasets. Convolutional neural networks (CNNs) were then trained to replicate aerothermal fields around a 1.0 m radius cylindrical model with a 1000 K wall temperature. The surrogate reproduced critical flow features, predicting maximum gas temperatures of 13,200–13,250 K near stagnation and peak pressures of 9.3–9.4 GPa, consistent with hypersonic re-entry conditions. Simulation runtimes were reduced from several hours (for full CFD) to less than 10 seconds, representing a computational saving of over three orders of magnitude, though with accuracy reduced by ~8–12% relative to wind tunnel and FEM benchmarks.

Building on this, a hybrid surrogate stress framework was developed to predict structural responses under hypersonic re-entry. Complex geometries were simplified into triangular and rectangular representations, then mapped into neural networks enhanced by low-rank approximation (retaining 16–64 dominant modes with <2% error). A dataset of 100 geometric variants, varying nose angle (15°–40°), radius (0.5–2.0 m), and body length (2–8 m), was combined with synthetic thermo-fluid and material datasets. Stress fields reconstructed via supervised regression and unsupervised clustering revealed ~16 distinct high-stress regions, with Von Mises stresses consistently reaching 380–420 MPa near stagnation and boundary layer separation zones. These predictions aligned within 5% of FEM and arc-jet test data for ZrB2–SiC composites.

Overall, this research demonstrates that ML-based surrogate frameworks can accurately reproduce degradation and stress distribution patterns while drastically reducing computational demands. Compared with full FEM analysis, surrogate modelling achieved up to 95% reductions in runtime with predictive errors below 10%. The approach provides a rapid exploratory tool for early-stage optimisation of UHTC materials and geometries. Remaining limitations include the treatment of non-linear temperature-dependent material properties above 2500 °C, dataset size restrictions, and partial loss of environmental fidelity.

In conclusion, the integration of ML surrogates with reduced-order CFD and stress clustering offers a scalable methodology for accelerating the design of non-ablative thermal protection systems. By combining computational efficiency with predictive reliability, this work supports faster material discovery cycles, improved identification of high-risk degradation zones, and robust design of components operating under hypersonic re-entry. Future research will focus on expanding datasets with high-fidelity wind tunnel data, refining non-linear constitutive models, and extending the surrogate framework to additional aerospace-relevant materials.