Hybrid Physics / Data-Driven Assembly Model

The problem

The assembly operation was governed by an analytical model based on nominal geometry and material behaviour. The model was correct in principle and consistently, slightly wrong in practice.

That is the interesting case. A residual that is systematic rather than random is not noise — it is structure the physical model does not describe. Operators were absorbing that structure manually, through re-measurement and rework loops. It worked. It cost time and parts.

Data

The measurements already existed, scattered across systems. Assembling them into a dataset that could be trusted took longer than building any model.

• Component geometry from inline metrology
• Process signals recorded during the assembly operation itself
• Context: tooling state, batch, machine, environment
• Outcome: the final measured result and the OK/NOK decision

The join was the hard part. Matching records by timestamp silently mixes parts; matching by serial requires the traceability to exist in the first place. Getting that right is the reason the model worked.

Approach

The key decision was not to replace the physical model, but to learn its residual.

   measurements ──→ physical model ──────────→ baseline prediction
         │                                            │
         └──────→ ML model on the residual ───────────┤
                                                      ▼
                                       corrected prediction + interval

Three reasons this formulation is better than a pure ML model here:

• Extrapolation degrades towards physics, not towards nonsense. Outside the training distribution the correction shrinks and the physical model carries the prediction.
• It is explainable to engineers. The learned part is a correction, and its feature importances map onto quantities the team already reasons about — so disagreements are about the process, not about the model.
• It fails gracefully. If the service is unavailable, the operation falls back to the analytical model, which is exactly what the line did before. A degraded day, not a stopped line.

I used AutoGluon to benchmark diverse architectures on the residual rather than committing to a favourite, then selected the top performer and verified it under a validation scheme grouped by batch and ordered in time. A naive random split looked dramatically better and was completely misleading, because parts from the same batch share tooling state.

Result

Scrap and rework rates in assembly operations dropped by 10 %. The more durable outcome is that the process became measurable: each rejected part now carries the feature that drove the rejection, which turned a recurring tooling issue — invisible before — into a scheduled maintenance rule.

What I would do differently

I would build the measurement database before proposing the model, not alongside it. Roughly two thirds of the elapsed time went into data plumbing, and the project plan did not say so at the start.