Bayesian Network for Uncertainty Quantification

The problem

A point prediction tells you what will probably happen. It does not tell you how confident to be, which parameter is responsible for the spread, or which operating point is robust rather than merely optimal on average.

For a process with many interacting parameters, those are the questions worth answering. An optimum sitting on a cliff edge is worse than a slightly lower one on a plateau — and a model that only outputs means cannot tell the difference.

Approach

I modelled the process as a probabilistic graphical model: nodes for the parameters and outcomes, edges for the dependencies between them, conditional distributions fitted from data where data existed and elicited from engineering knowledge where it did not.

   tooling ─┐
            ├─→ process state ─┬─→ geometry ─→ outcome
   material ┘                  └─→ energy input ─┘

That structure buys three things a black-box regressor does not give you:

1. Dependencies are explicit and inspectable. The graph is a hypothesis about the process, drawn in a form the engineers can argue with. Several edges were removed in exactly that way.
2. You can condition on anything. Fix a parameter, observe an outcome, propagate — the same model answers forward and inverse questions.
3. Uncertainty propagates natively. It is not bolted on afterwards.

Monte Carlo simulation over the network then produced full outcome distributions rather than point estimates, which is what made the uncertainty quantification rigorous instead of decorative: sensitivity attributed to each parameter, tail behaviour made visible, and robustness of a candidate operating point measurable.

Result

The optimisation changed character. Instead of “this setting maximises the expected outcome”, the answer became “this setting keeps the outcome inside specification across the realistic range of variation, and here is which parameter you should control more tightly to shrink the spread further.”

That is a recommendation a process owner can act on, and one they can challenge — which matters more than being right on average.