Conclusion

All physics-informed models outperform the baseline RNN and NODE-ED in both R² and MSE. The best performance is achieved by NODE-ED-Uncoupled (SS-C-PD), with improvements of 6.15% and 9.87% over RNN, and 5.03% and 12.15% over NODE-ED in R² and MSE respectively. Additionally, the highest reported R² and MSE values for NODE-ED-Uncoupled Bagging are at 0.773 ± 0.0882 and 0.239 ± 0.070 respectively. Notably, architectural modifications alone also yield improvements, with NODE-ED-Uncoupled outperforming NODE-ED by 3.75% in R² and 7.54% in MSE.

While the results in this work showed only a moderate improvement in modelling accuracy, even a small gain can lead to substantial savings. In a sensitivity analysis of chemisorption CO₂ capture cost, Nuchitprasittichai and Cremaschi (2013)⁶¹ reported that a ± 50% change in CO₂ concentration at the tower inlet had a strong effect on cost. Specifically, a 50% decrease in CO₂ concentration increased capture cost by 85.8%, and this increase is estimated to be about $128,000/day in an industrial setting⁶². Additionally, Vinjarapu et al. (2024)⁶³ reported that in Saga City, Japan, a typical flow rate of captured CO₂ waste gas is 10,000 kg/day. A 1% prediction error in CO₂ therefore corresponds to an additional release of 100 kg/day of CO₂ to the environment. As such, even though this work showed a moderate improvement, it should be acknowledged that in chemical engineering settings these improvements are substantial both economically and environmentally.

Importantly, the value of the developed PIML models extends beyond immediate performance gains, they improve generalisability and ensure that predictions remain consistent with governing principles under unseen conditions. As such despite a modest improvement and larger complexity, the enhanced physical interpretability and model trustworthiness is valuable especially in the field of chemical engineering where predictions are heavily grounded in physical phenomenon.

1. 61 Nuchitprasittichai, A.; Cremaschi, S. Sensitivity of Amine-Based CO₂ Capture Cost: The Influences of CO₂ Concentration in Flue Gas and Utility Cost Fluctuations. International Journal of Greenhouse Gas Control 2013, 13, 34-43. https://doi.org/10.1016/j.ijggc.2012.12.012. ↩
2. 62 Putta, K. R.; Saldana, D.; Campbell, M.; Shah, M. I. Development of CO₂ Capture Process Cost Baseline for 555 MWe NGCC Power Plant Using Standard MEA Solution. SSRN Electronic Journal 2022. https://doi.org/10.2139/ssrn.4279671. ↩
3. 63 Vinjarapu, S. H. B.; Neerup, R.; Hellerup Larsen, A.; Nis Bay Villadsen, S.; von Solms, N.; Jensen, S.; Lindkvist Karlsson, J.; Kappel, J.; Lassen, H.; Blinksbjerg, P.; Loldrup Fosbol, P. Pilot-Scale CO₂ Capture Demonstration of Heat Integration through Split Flow Configuration Using 30 wt% MEA at a Waste-to-Energy Facility. https://www.sciencedirect.com/science/article/pii/S1383586624010505#b6 (accessed 2026-04-01). ↩