A physics-guided fourier-transformer surrogate model for a vapor-compression cycle with strong generalization under distribution shift: Benchmarking against koopman-inspired state-space learning
International Journal of Heat and Fluid Flow, cilt.122, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 122
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.ijheatfluidflow.2026.110764
- Dergi Adı: International Journal of Heat and Fluid Flow
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Applied Science & Technology Source, Chimica, Compendex, INSPEC, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO)
- Anahtar Kelimeler: Distribution shift, Fourier features, Koopman state-space model, Physics-guided learning, Thermodynamics assessment, Transformer encoder, Vapor-compression cycle
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
Accurate modelling of the performance of a vapor-compression cycle across broad operating ranges is essential for the design, optimization and supervisory control of refrigeration systems. But traditional data-driven surrogate models can perform poorly out of training distribution and do not guarantee thermodynamic feasibility. Here we introduce a physics-guided Fourier-transformer surrogate model to model a single-stage vapor-compression cycle with R1234yf and nano-lubricant additives, such as alumina, graphene, and carbon nanotubes. The model leverages Fourier feature lifting for smoother non-linear interactions, Transformer blocks to learn interactions across the four thermodynamic states, and physics-guided soft penalties designed to encourage physically plausible predictions by penalizing violations of output bounds and a prescribed monotonicity prior, without explicitly enforcing the governing thermodynamic conservation equations. To facilitate a systematic evaluation, a Koopman-inspired physics-guided state-space model is also proposed and compared with ANN, PIML, PGML, and PG-FNT models. The novelty of this study lies in its combination of spectral, relational, and physics-guided learning within a single compact architecture and its comparison with a specially designed state-space model under explicit distribution shift testing. To the authors’ knowledge, this combination has not been previously investigated for vapor compression refrigeration systems using nanolubricants with low GWP refrigerants. In the original five-fold cross-validation benchmark, PG-FNT[full] achieved a mean validation RMSE of 0.005202 and MAE of 0.003895, corresponding to reductions of approximately 55.5% and 55.7%, respectively, compared with the conventional ANN baseline under the same evaluation protocol. PG-KSSM also demonstrated competitive predictive performance, achieving up to 28.7% lower RMSE than ANN in the corresponding benchmark. Furthermore, under a high-condenser-temperature operating-envelope shift within the same experimental campaign, PG-FNT achieved a mean R2 of 0.9541 ± 0.0082 across five initialization seeds following few-shot adaptation with 10 labelled observations. Performance was evaluated on a separate 20-sample held-out test set, with no overlap between adaptation and test observations. These findings demonstrate the strong predictive capabilities of the proposed architectures and highlight the effectiveness of limited-data adaptation for maintaining accurate thermodynamic performance prediction under changing operating conditions. This work presents a transferable benchmarking and cross-validation, out-of-dispersion assessment, perturbation analysis, and significance testing protocol applicable to other thermofluidic systems with limited small and noisy datasets for the scientific community. The proposed framework provides a physics-guided, data-driven approach for predicting vapor-compression-cycle performance, demonstrating improved accuracy relative to the evaluated machine-learning baselines. Its comparative performance against conventional thermodynamic cycle solvers remains to be established. Looking ahead, this framework provides a foundation for fault-tolerant digital twins and real-time supervisory control, while the competitive performance of PG-KSSM points to reduced-order, control-oriented thermodynamic models for future closed-loop applications.