Abstract
<title>Abstract</title> <p>Lithium-ion battery management systems infer internal safety states mainly from surface temperature, voltage and current, leaving the coupled thermal and mechanical processes inside cells only indirectly observable. Embedded fibre Bragg grating sensors can access these states in situ, but a single wavelength shift combines temperature and strain into one measurement, creating an underdetermined inverse problem. Here we show that this single-fibre observability problem can be constrained by battery physics and solved in real time. We introduce FBG-PhysNet, a dual-branch recurrent framework that combines fixed-state RWKV-7 inference, electrochemical heat-generation conditioning and nine differentiable constraints spanning Bragg optics, heat conduction, energy conservation, mechanical feasibility and electrochemical consistency. Across 75,120 hardware-in-the-loop windows from five operating regimes, the model estimates internal temperature and strain with 2.270 K and 21.94 µε mean absolute errors, reduces the cross-talk ratio to 58.6%, and retains a fixed recurrent state of 64 KB per RWKV block. These results establish a deployable route to internal thermo-mechanical observability for adaptive battery safety management.</p>