Bridging lab to industry: Deep learning cuts testing time for lithium metal battery health models

Reliable state of health estimation is vital for lithium metal batteries, yet existing approaches are constrained by high data demands and costly testing, limiting their scalability from laboratory to industry. Here, we present a transferable state of health modeling framework based on a stacking autoencoder-fully connected network. By integrating ensemble learning with hierarchical autoencoder representations, the framework automatically derives health indicators from as few as 21 or even 6 relaxation voltage points, eliminating the need for handcrafted features. Coupled with a lightweight predictor, it achieves accurate state of health estimation with a mean absolute error of 0.56 % and a root mean square error of 0.74 %. Beyond accuracy, the framework demonstrates strong adaptability across capacity scales and inter-manufacturer formulation-process variations: when transferring from 3.8 mAh laboratory cells to 19.9 Ah industrial batteries, estimation errors are reduced by up to 67.12 % (Mean Absolute Error) and 63.52 % (Root Mean Squared Error). This work provides a scalable and cost-effective pathway for rapid state of health model deployment, supporting efficient and safe industrial application of lithium metal batteries.