Interpretable domain adaptation enables robust lactic acid fermentation monitoring from waste

• Transferring predictive models without chemical analysis from the new environment. • Addressing FTIR shifts using DANN-R, DeepCORAL-R, DARE-GRAM, and TCA methods. • Improving transparency and reliability in glucose and lactic acid predictions. • Enhancing monitoring accuracy with ShapDA by identifying domain-invariant features. Lactic acid is a key compound in the circular bioeconomy, supporting the production of biodegradable plastics, pharmaceuticals, and food products. To optimize its production, real-time fermentation monitoring is essential, particularly when using heterogeneous substrates such as agricultural residues or municipal bio-waste. While manual sampling and offline chemical analysis are impractical at scale, machine learning models coupled with spectroscopic sensors offer a promising alternative. However, they often fail to generalize across varying substrates (e.g., glucose and bio-waste) due to data distribution shifts, limiting industrial deployment. To address this, we present Shapley Additive exPlanations-based Domain Adaptation (ShapDA), an unsupervised domain adaptation framework that integrates model explainability with Fourier transform infrared spectroscopy. Unlike conventional domain adaptation methods that align statistical distributions or complex latent representations, ShapDA identifies a compact set of spectral features that remain stable across substrate domains. These domain-invariant features capture variable-specific spectral signatures, enabling calibration-free prediction under novel conditions. We apply ShapDA to 21 glucose-based fermentation batches, each consisting of spectra and their corresponding chemical measurements. We then transfer it to 83 batches of bio-waste fermentation-spanning diverse substrates, production scales, and microbial strains-without additional chemical measurements. ShapDA outperforms baseline and state-of-the-art methods, reducing the root mean squared error by 50 % for glucose and 65 % for lactic acid in bio-waste fermentations. These results demonstrate that ShapDA is a scalable solution for substrate-independent monitoring across diverse production environments. Our code is available at https://github.com/shl-shawn/ShapDA .