Heriot-Watt Wants AI to Set Safety Boundaries for Underground Energy Projects
A new perspective paper argues the real bottleneck in subsurface ML is not accuracy but calibrated uncertainty and auditability for regulators.
- Perspective paper reframes subsurface ML: deliverable is calibrated uncertainty and auditable risk envelopes, not point forecasts.
- Identifies four bottlenecks: scarce labels, uncorrelated uncertainty, poor scale bridging, weak QA for regulators.
- Recommends hybrid physics-ML, ensemble inversion (ES-MDA/EnKF) accelerated by surrogates, and structure-aware generative priors.
- Positions foundation models narrowly: converting unstructured legacy reports into checkable structured priors, not autonomous decisions.
- Four anchor applications: pore-scale digital twins, near-well multiphase, MMV inverse problems, basin-scale portfolios.
- Calls for 2-3 open reference testbeds by 2028 with fixed evaluation protocols and microseismicity data.
A new perspective paper from researchers at Heriot-Watt University and Subsurface Labs argues that machine-learning systems for geoenergy should produce calibrated, risk-bounded operating envelopes. These envelopes define ranges of injection rates, pressures, and other controls that meet performance and safety constraints at a stated confidence level, giving operators actionable limits and regulators an auditable evidence trail.
The proposal covers carbon dioxide storage, geothermal energy, underground hydrogen storage, and critical-mineral extraction. These projects have long operating lives and limited field data, so model validation often lags deployment. Forecast errors can contribute to permit delays, excessive pressure, induced seismicity, poor well performance, or carbon dioxide leakage.
Sparse data, slow ground truth
Geoenergy strains common machine-learning assumptions because labels are scarce, measurements are noisy, and several geological models can explain the same observations. This non-uniqueness complicates inverse problems, which infer hidden subsurface properties from seismic surveys, well logs, pressure readings, and deformation measurements. The governing physics is also incomplete at relevant scales, while definitive outcomes may take decades to observe.
Four blockers to deployment
- Scarce and biased labels: Field outcomes arrive slowly and contain operational confounders. Laboratory and synthetic datasets may transfer poorly to new sites, while instruments and measurement practices change over time.
- Incomplete uncertainty estimates: A point forecast cannot show whether a proposed operating plan remains safe across plausible geological conditions.
- Weak scale bridging: Models must connect micron-scale pore structure with reservoir and basin behavior across hundreds of kilometers, including coupled flow, chemistry, heat transfer, and geomechanics.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
