Screening one candidate injection site the traditional way takes roughly three weeks of specialist review. In a synthetic demonstration, the same screen ran in under a minute, which is about forty sites a day from one workstation. Every number in this piece is illustrative and comes from that demonstration, not from a deployment.
The problem is throughput, not accuracy
The usual framing is that screening should be more accurate. That is rarely the binding constraint. Specialist geomechanical review is already good; there is simply not enough of it, and its cost is paid per site.
That per-site cost has a consequence nobody plans for. When a screen costs three weeks, a team screens the sites it already believes in. The portfolio never gets an even look, and the sites that would have surprised you are exactly the ones that never get the budget. The bias is invisible because it lives in what was never assessed.
What amortization changes
Amortized inference moves the expensive computation from per-site to once. Instead of solving the geomechanical problem for each candidate, you train a surrogate on many solved instances, and afterwards each new site is a forward pass. The physics has not been skipped; it has been paid for in advance and reused.
The economics invert. Under per-site cost, screening more sites costs proportionally more, so you screen fewer. Under amortized cost, the marginal site is nearly free, so the rational move is to screen everything, including the sites you expect to reject. Rejecting cheaply is itself valuable, because it is evidence rather than assumption.
Reading the screen, and what it does not do
The screen combines injection rate, the stress margin to failure, and distance to the nearest mapped fault into a score, and returns a recommended rate cap. The interaction is the part worth understanding: no single input determines the verdict.
Push the injection rate up when the stress margin is comfortable and the nearest fault is far, and the cap gives ground. Push the same rate with a fault nearby and it stops giving ground almost entirely. That refusal is the model behaving correctly, and it is the most useful thing it does in a negotiation, because it makes the trade explicit rather than leaving it to be argued.
Traditional screen, per site
Amortized screen, per site
Sites per day, one workstation
Inputs combined into the verdict
How a decision-maker should read it
Three cautions belong with any number this screen produces.
It is a screen, not a clearance. Its output ranks and caps candidates so that specialist review is spent where it matters; it does not replace that review, and a site the screen passes still needs one.
Its confidence is bounded by its training distribution. A site geologically unlike anything the surrogate saw should be escalated rather than scored, which requires the surrogate to report uncertainty rather than only a number.
And a recommended cap is a starting condition, not a permit. Real injection is monitored and adjusted; the screen sets where to begin and what to watch.
Zooming out
The pattern generalises well past induced seismicity. Wherever a rigorous assessment is too expensive to run at portfolio scale, the outcome is not that decisions wait for the assessment. It is that most decisions get made without it, on judgement, and only the shortlist gets the rigour. Amortizing the assessment does not make it more accurate. It makes it available to every candidate, which changes which candidates get considered at all.
Key takeaways
- The binding constraint in site screening is throughput rather than accuracy: specialist review is good and there is not enough of it.
- Per-site cost biases portfolios invisibly, because teams screen the sites they already believe in and the surprising ones are never assessed.
- Amortization moves the expensive computation from per-site to once, so the marginal site becomes nearly free and screening everything becomes rational.
- The verdict comes from the interaction of rate, stress margin and fault distance, not from any single input, which makes the trade explicit in a negotiation.
- A screen ranks and caps so specialist review is spent well; it is not a clearance, and a site unlike the training distribution should be escalated rather than scored.
