Multipath Adaptive Gated Bottleneck Latent ODE with Raman Data Fusion for Cell Culture Process Forecasting

Biopharmaceutical manufacturing, particularly mammalian cell culture, suffers from high variability where runs with similar early-stage behavior can diverge into drastically different outcomes. Because these processes are expensive and slow, and measurements are often sparse or irregularly sampled, traditional forecasting models struggle to provide accurate, actionable predictions. This paper addresses the challenge of predicting multi-day future trajectories in heterogeneous fed-batch bioreactor runs.

The authors propose a two-part framework: the Gated Bottleneck Latent ODE (GB-Latent ODE) and Multi-Path Just-In-Time Fine-Tuning (MP-JIT-FT). The GB-Latent ODE is a neural ordinary differential equation model designed to handle sparse, high-dimensional data by using learnable variable-wise gating and a bottleneck layer to compress inputs.

To address the problem of multiple possible futures, the MP-JIT-FT component retrieves historical runs similar to the current process, clusters them into distinct regimes, and fine-tunes separate model copies for each regime. This allows the system to output multiple plausible future paths rather than a single, potentially misleading average. Additionally, the authors integrate Raman spectroscopy data via a soft sensor, which provides dense pseudo-observations to enrich the sparse offline measurements.

Evaluated on 38 fed-batch bioreactor runs across 14 conditions, the proposed framework achieved the best average rank compared to a global Latent ODE baseline. The study demonstrates that the multi-path approach is particularly effective when early-stage dynamics are similar but lead to divergent futures. Furthermore, the integration of Raman data significantly improves performance when early process dynamics are representative of later behavior, providing a more robust signal for mid-run adaptation.