A company with six plants running the same reaction has, in principle, six times the calibration data of a company with one. In practice it rarely gets to use them. The spectra sit behind separate site firewalls, sometimes behind separate legal entities after an acquisition, and in contract manufacturing they belong to different customers who have no interest in seeing each other’s process signatures. The standard workaround is to build one model per site and accept that each is trained on a narrow slice of the variation the process can actually produce.
Federated learning proposes a different arrangement. Each site trains on its own data and ships only model parameters to an aggregator, which averages them into a global model and sends it back. Raw spectra never leave the site. The idea has been established in medical imaging and drug discovery for several years. What changed over the past twelve months is that it started appearing in the chemometrics and chemical-engineering literature with spectroscopic data attached rather than as a general proposal.
This roundup looks at what those papers actually demonstrated, and at the gap between the machine-learning result and what a validated production model would need.
The framing papers
Two 2025 publications set out the case for the process industries rather than for hospitals. Dutta and co-authors published a tutorial in Industrial & Engineering Chemistry Research (volume 64, pages 7767-7783) that walks through federated implementations in the Flower and TensorFlow Federated frameworks, with worked chemical-engineering examples. Their stated finding is that federation often maintains or improves classification performance relative to centralised training, and that the advantage is largest where the client datasets are heterogeneous rather than near-identical.
That last point matters more than it sounds. If six sites hold statistically similar data, federation buys little beyond compliance convenience. It pays where sites differ - different feedstock suppliers, different reactor geometries, different seasonal ranges - because the global model sees variation no single site could contribute.
Rittig and Kortmann’s perspective piece (arXiv:2506.18525, June 2025) frames the problem in blunter commercial terms: chemical and process data are proprietary and therefore locked in silos, and the models the industry could build are constrained by that fact rather than by algorithmic limits. They demonstrate the argument on two cases, activity-coefficient prediction with graph neural networks and distillation-column identification with autoencoders, reporting that federated models outperform models trained by individual companies while approaching the accuracy of a centralised model built on the pooled set.
Spectroscopy-specific results
The first paper to put NIR calibration itself at the centre is Chen’s proof-of-concept study in Chemometrics and Intelligent Laboratory Systems (volume 268, article 105583, January 2026), which tests a decentralised federated architecture - no central aggregation server - for collaborative NIR calibration across institutions that do not share raw spectra. The framing is deliberately modest: it is a proof of concept, and the authors present it as evidence that privacy-preserving calibration is feasible rather than as a deployable protocol.
A second NIR application, from Chen and colleagues in Computers and Electronics in Agriculture (volume 239, article 111085, December 2025), applies a federated framework built on deep echo-state networks to soil organic carbon prediction. The agricultural setting is not incidental. Soil spectral libraries are the clearest existing case of many parties holding comparable NIR data that they will not merge, which makes the domain a natural first testbed for methods that industrial users will eventually want.
On the process side, an April 2026 preprint by Pipattaratonchai and Aueawatthanaphisut (arXiv:2604.26073) applies secure aggregation to neural process models trained on plant sensor data, reporting that the global mean squared error fell from roughly 2369 to below 50 within five communication rounds. The convergence behaviour is the useful observation - not the absolute error, which is specific to their simulated setup.
The problem federation does not solve
None of this addresses the reason multi-site spectroscopic models fail in practice, which is that the instruments differ. Two probes of the same model number, on the same reaction, produce spectra with systematic offsets in wavelength axis, intensity response, and baseline. Averaging model weights across sites does not remove instrument bias; it averages it in.
That is a calibration-transfer problem, and it has its own active literature. Lai and co-authors, writing in Analytical Chemistry (volume 97, issue 35, pages 19009-19018, 2025), applied low-rank adaptation to transfer deep-learning Raman models between spectrometers, reducing trainable parameters by a factor of 600 against full fine-tuning. On their methanol mixture test set the method reached R2 = 0.952 with RMSE 0.072, against 0.846 and 0.117 for piecewise direct standardisation and 0.863 and 0.114 for full-parameter fine-tuning.
The architectural similarity is worth noting. Low-rank adaptation keeps a shared backbone and attaches a small per-instrument module. That is close to what a federated deployment with heterogeneous hardware would need: a global model carrying the chemistry, plus a light local correction carrying the instrument. No paper this year has combined the two, and doing so is the obvious next step. Readers working on the instrument side of this will find the wider treatment in our review of calibration transfer between instruments more directly applicable today.
What a regulated deployment would still require
The literature is a machine-learning literature. It reports accuracy against held-out sets and says almost nothing about the artefacts a GMP model needs. A federated model has no single training dataset to file, which complicates the model-description sections that regulators expect; the aggregation history becomes part of the model’s provenance, and there is currently no established convention for documenting it. Change control is harder still - every aggregation round is arguably a model update, and the classification of post-approval changes assumes updates are discrete events with an identifiable trigger.
Drift monitoring also has to work differently. When a global model degrades at one site, the ordinary diagnostic question - has the process moved or has the model gone stale - gains a third possibility, which is that another participant’s data pulled the shared weights away. The drift detection routines that work for single-site models were not designed to attribute degradation across participants.
For teams evaluating whether any of this is near-term, the practical answer is that it is not. The organisational work of a multi-site model rollout - agreeing reference methods, harmonising sampling, standardising instruments - has to be done first, and it is the same work whether the final model is federated or centralised. Federation removes the last obstacle, not the first ones.
What the 2026 papers establish is narrower but real: the accuracy penalty for not pooling spectra is smaller than the field assumed, and the case for keeping data siloed can no longer rest on the claim that federated models do not work. The remaining barriers are instrumental and procedural, and those are the ones worth working on now.