TCPFN Foundation Model for Industry 4.0

TCPFN is a transformer-based causal foundation model that answers "why" questions on industrial time-series data. Given raw multivariate sensor data from a plant historian, TCPFN infers the temporal causal graph connecting variables, traces observed events back to their root causes, and estimates the effect of candidate interventions. It does all of this zero-shot: no site-specific training, no labeled failure data, and no feature engineering.

Most industrial AI models are correlational. They can tell you that a quality deviation and a pressure drift move together, but not which one drives the other, or what will happen if an operator intervenes. TCPFN is built for the causal questions that determine action: what caused this event, what should we change, and what will happen if we do.

TCPFN follows the prior-data fitted network (PFN) approach introduced by Muller et al. (arXiv:2112.10510) and popularized by TabPFN (Hollmann et al., Nature 2025). Instead of training on customer data, the model is pretrained once on millions of synthetic systems sampled from a prior over temporal structural causal models. At inference time it conditions on the customer's observed data in-context and outputs an amortized approximation of the Bayesian posterior over causal structure. Customer data is never used for training and never leaves the inference boundary.

Model architecture

• Transformer encoder, 21.6M parameters
• Pretrained for 200K steps on 30 million synthetic temporal structural causal systems
• In-context Bayesian inference: the model approximates the posterior over causal graphs and effects conditioned on the provided dataset
• Reference: Temporal Causal Prior-Data Fitted Networks, arXiv:2606.20889