Physics Informed Constrained Reinforcement Learning for Carbon-Regulated Cold Chain Inventory with Digital Twins

Authors

  • Ashish Kumar Routray Ravenshaw University Author
  • Swagatika Dehury Ravenshaw University Author
  • Mahottak Kumar Synergy Polytechnic Author
  • Sudhir Kumar Sahu Ravenshaw University Author

DOI:

https://doi.org/10.67014/jadeia.v1i1.0030

Keywords:

AI empowered engineering, physics informed neural networks, constrained reinforcement learning, cold chain logistics, non-instantaneous deterioration, carbon cap and trade, synthetic simulation proof of concept

Abstract

Refrigerated warehousing and reefer distribution represent some of the most energy intensive and loss prone links in critical food and pharmaceutical infrastructure, yet operational control remains dominated by static thermal setpoints and uncoupled lot-sizing heuristics. To address this limitation, we investigate PI-CSAC, a physics informed constrained actor critic framework designed to jointly optimize replenishment lot sizes, dynamic chamber temperature setpoints, and carbon allowance trading actions under cap and trade regulation. The formulation couples an Arrhenius constrained neural deterioration surrogate with a constrained Markov decision process solved via online Lagrangian dual relaxation against explicit fill rate and emissions constraints. In a synthetic proof of concept benchmark using literature derived Arrhenius kinetics across three illustrative SKU families evaluated over twenty held-out years, a reduced scale constrained agent (implemented with REINFORCE, a learned value baseline, and Lagrangian dual ascent) was compared against a grid optimized closed form non-instantaneous deterioration EOQ policy and a grid optimized (s,S) heuristic with safety stock. The reduced scale agent did not outperform either analytical baseline, resulting in 7% to 28% higher mean annual costs across families alongside higher emissions variance and no clear performance separation in constrained versus unconstrained ablations, despite delivering modestly higher fill rates. Ultimately, these findings deliver critical negative and mixed empirical evidence establishing that computational tractability does not guarantee immediate field deployability, highlighting that joint thermo-inventory reinforcement learning requires substantially greater network capacity, training scale, and domain calibration before surpassing tuned classical baselines.

Downloads

Published

September 28, 2026

How to Cite

[1]
A. K. Routray, S. Dehury, M. Kumar, and S. K. Sahu, “Physics Informed Constrained Reinforcement Learning for Carbon-Regulated Cold Chain Inventory with Digital Twins”, JADEIA, vol. 1, no. 1, p. 21, Sep. 2026, doi: 10.67014/jadeia.v1i1.0030.