Conservation-informed machine learning for physically admissible four-phase hydrothermal liquefaction surrogate modelling.

Hydrothermal liquefaction (HTL) distributes wet biomass and residues among bio-oil, char, aqueous product, and gas. Independent phase regressors can produce negative or non-closed allocations, and random validation obscures inter-study heterogeneity. We develop a conservation-informed constrained surrogate treating the four yields as one composition vector. A curated corpus of 3693 records yielded 680 closure-valid HTL and solvothermal experiments from 60 publications and 12 feedstock families. A Level 0-5 hierarchy used a common 408/136/136 training/calibration/test partition. The selected L5 surrogate combines an XGBoost teacher with a softmax admissibility mapping and soft conservation-residual penalties. On the 136-record test set, L5 reached a mean four-phase R2 of 0.844 and a vector root-mean-square error (RMSE) of 6.462 wt%, a small paired difference from the strongest comparator (ΔR2=-0.0059; 95% confidence interval, -0.013 to ＋0.001), while enforcing non-negativity and exact closure and reducing common-basis carbon and higher heating value residuals to 2.390 wt% and 1.028. Under 10% feature perturbation, L5 produced 0% closure failures and 4.02% carbon-balance threshold violations, versus 50.28% and 69.82% for the independent baseline. Publication-aware validation exposed inter-study heterogeneity. Among 38 publications eligible for five-record local calibration, median vector RMSE decreased from 13.91 (k=0) to 7.10 wt% (k=5). Literature-derived HTL surrogates should be evaluated beyond random cross-validation and locally calibrated before transfer. Phase-wise 90% split-conformal coverage ranged from 0.875 to 0.956. The framework supports uncertainty-aware, physically admissible screening within represented data support and calibration-informed deployment, not universal prediction of unseen studies or full process-economic and life-cycle modelling.