Speaker
Description
Accurate aerothermodynamic modelling of space debris reentry is critical both for ground
casualty risk assessment and for quantifying ablation-driven mass deposition in the
upper atmosphere. Operational tools, such as ESA’s DRAMA software suite,
rely on engineering correlations with simplified chemical assumptions, while high-fidelity
CFD, though resolving some of these limitations, remains prohibitive for the probabilistic
trajectory ensembles required in Design-for-Demise workflows. Moreover, part of the
chemistry occurring during reentry remains difficult to capture even at high fidelity. This
motivates the development of fast, physically grounded surrogate models that can serve
as a foundation for progressively more complex coupled aerothermochemical problems.
A 0D reference solver is constructed to generate the training database. Post-shock
conditions are obtained from the Rankine–Hugoniot relations solved iteratively via
Newton–Raphson, with thermochemical equilibrium computed through Gibbs free-
energy minimisation using the Mutation++ thermochemical library. Wall boundary
conditions, for a carbon-based spherical object, are handled through a Surface Mass
Balance (SMB) formulation accounting for carbon injection via sublimation (C3
formation), oxidation, and oxygen catalysis. This yields mass blowing rate, heat flux
partition, and near-wall species mass fractions. The input space spans altitude 30–80 km,
Mach 20–28, and body radius 0.1–1.0 m, covering the continuum reentry corridor for low-
Earth-orbit debris.
Three tree-based ensemble regressors (Decision Tree, Random Forest, and XGBoost) are
benchmarked in a unified multi-output architecture predicting eleven quantities
simultaneously: total heat flux, mass blowing rate, electron number density, and nine
near-wall species mass fractions. Hyperparameter optimisation is performed via Optuna,
a Bayesian framework based on the Tree-structured Parzen Estimator, with 3-fold cross-
validation and automated median pruning to ensure that observed performance
differences are attributable to algorithmic structure rather than tuning disparity.
XGBoost achieves R²scaled = 0.999976 on the test set, outperforming Random Forest
(0.999935) and Decision Tree (0.999720). Optuna-tuned XGBoost reduces MAEscaled by
a factor of three relative to a default baseline, with heat flux identified as the hardest
output through a composite difficulty index combining R², MAE and RMSE across all
targets.
These results establish tree-based ensembles as a practical and competitive algorithmic
family for aerothermochemical surrogate modelling, currently underexplored relative to
ANNs and kriging in the debris reentry literature. The unified multi-output architecture,
providing heat flux, blowing rate, and ablation product composition simultaneously, is
directly amenable to embedding in higher-fidelity reentry codes.
This work is conducted within the ESA OSIP PhD fellowship framework, which aims to
predict element-specific atmospheric mass deposition profiles (such as Al, Cu, or C3 ).
The present surrogate constitutes the first building block of that chain. Validation of the
modelling approach against plasma wind tunnel data (VKI Plasmatron) and in-flight
optical observations is planned as next steps, targeting a first operationally viable
environmental footprint assessment within a reentry modelling chain.