Aleatoric versus Epistemic Uncertainty
Aleatoric uncertainty is inherent randomness you cannot reduce; epistemic uncertainty is ignorance you can reduce with more data or better models.
Two fundamentally different kinds
Uncertainty splits into two categories that behave differently and demand different responses. Aleatoric uncertainty (from the Latin for dice) is the irreducible variability of a system - genuine randomness that persists no matter how much you learn. Epistemic uncertainty (from the Greek for knowledge) is due to limited information and shrinks as you gather more data or improve the model.
Examples
- Aleatoric: thermal noise in a sensor, turbulent fluctuations, manufacturing scatter across nominally identical parts
- Epistemic: an imprecisely known reaction rate, an untested model assumption, a parameter fit from too few samples
Why the distinction matters
The two call for opposite strategies. You cannot buy down aleatoric uncertainty with more experiments - only by changing the system or accepting the spread. Epistemic uncertainty is a to-do list: more measurements, better physics, or more training samples will reduce it. Confusing the two leads to wasted effort on the irreducible or false confidence about the reducible.
How they appear in models
In a probabilistic model, aleatoric uncertainty is often modeled as observation noise - a variance term on the output. Epistemic uncertainty appears as uncertainty in parameters or model structure - a distribution over weights, a Gaussian-process posterior variance, or the spread across an ensemble. A Gaussian process, for instance, separates its predictive variance into a noise floor (aleatoric) and a distance-dependent term that vanishes near data (epistemic).
Combining them
Total predictive uncertainty is the combination of both. Deep ensembles and Bayesian neural networks that predict both a mean and a variance can decompose their output uncertainty into the two parts, which is valuable for deciding whether more data would help.
In fusion UQ
For Kronos models, plasma turbulence and shot-to-shot variability are aleatoric, while imperfectly known transport coefficients and model-form choices are epistemic. Separating them clarifies which uncertainties in the machines' predicted performance more experiments can tighten and which reflect the physics itself.