By Saad Iqbal
A platform has been injecting seawater to hold reservoir pressure for a decade without issue. Then, one quarter, the H2S reading at the gas plant inlet starts climbing — not from a leak, not from a new zone, but from chemistry that was set in motion years earlier, deep in the reservoir, the moment seawater sulfate met the right bacteria and the right food source. By the time anyone sees it at surface, the souring is already years into a process nobody can simply turn off. That lag is what makes reservoir souring one of the few upstream problems where the whole value of prediction is in how early you look, not how accurately you react.
What Actually Sours a Reservoir
Seawater carries sulfate. Once it’s injected, sulfate-reducing bacteria (SRB) — whether native to the formation or carried in with the injection water — consume that sulfate along with available nutrients, mainly volatile fatty acids (VFAs) in the produced and connate water, and excrete hydrogen sulfide as a metabolic byproduct. The gas is toxic, corrosive to wellbore tubulars and surface facilities, and expensive to handle once it reaches surface. Souring typically shows up at a producing well only after water breakthrough, and tends to worsen as water cut climbs — which is exactly why it’s so easy to miss during the planning phase, when the well is still producing clean and dry.
A Real Case That Shows Why This Is Genuinely Hard to Predict
An SPE study of two Middle Eastern fields (identified only as Field S and Field T) built a compositional, non-isothermal dynamic model — three phases, eighteen components, eighteen reactions — to forecast most-likely and worst-case H2S levels across field life. The spread between those two cases is the real story. In Field S, which already carried 350 ppmv of background (non-microbial) H2S, the base case had microbial souring pushing that to roughly 1,000 ppmv after water breakthrough; the worst case, built on underestimating VFA content and overestimating the reservoir’s natural H2S abstraction capacity, reached about 3,400 ppmv. In Field T, which started with no background H2S at all, the base case reached 195 ppmv by 2044 at 95% water cut, while the worst case climbed to roughly 1,200 ppmv.
That three-to-sixfold swing between base and worst case isn’t modeling sloppiness — it’s the honest range you get when two genuinely uncertain inputs (nutrient availability and the rock’s own capacity to re-absorb H2S) are allowed to vary realistically. A single deterministic forecast tells you almost nothing useful here; you need the distribution, not the point estimate.
How Operators Actually Monitor for Souring
Waiting for a rising H2S reading at the separator is the lagging indicator — by then the microbial community is already established. The leading indicators sit upstream, in the produced and injection water itself. Culture-based most-probable-number (MPN) counts have historically been the field standard for enumerating SRB, but they undercount badly because most souring-relevant organisms grow poorly, or not at all, outside the reservoir; quantitative PCR (qPCR), which counts gene copies directly from a water sample rather than waiting on bacterial growth, has become the more reliable molecular alternative for tracking SRB population trends over time. Alongside biological counts, operators track redox potential (ORP) in produced water — a sharply negative swing signals the reducing conditions SRB need — and watch for iron sulfide scale showing up in separators and pigging runs, itself a byproduct of H2S reacting with dissolved iron and a sign the reaction is already well underway downstream. None of these indicators predicts souring on their own the way a compositional model does; together, they tell you whether the mechanistic forecast is tracking reality or needs re-calibrating against what the field is actually doing.
The Nitrate Trade-off Nobody Expects
The standard mitigation is injecting nitrate or nitrite alongside the seawater: it feeds nitrate-reducing bacteria that outcompete SRB for the same nutrient pool, in theory starving the souring reaction before it starts. Researchers at DTU modeling this at field scale found something counterintuitive: under-dosed nitrate can lower the total H2S generated across the reservoir while simultaneously shifting where it’s generated — closer to the producing wells. A treatment program can look like a clear win on a reservoir-wide mass balance and still deliver more H2S to the wellhead than doing nothing, because where the reaction happens matters as much as how much reaction happens.
In the Field S/T study, the most robust prevention option wasn’t nitrate at all — it was re-injecting the field’s own high-salinity produced water (already chemically equilibrated and poor SRB habitat) instead of fresh low-salinity aquifer water. It kept H2S near background levels even in the worst case. The catch: it requires produced-water transfer infrastructure that, for both fields in the study, simply didn’t exist yet.
Where AI Actually Fits In
Let’s be precise about what’s doing the predicting here, because it’s not machine learning: it’s mechanistic chemistry — a compositional, reactive-transport simulation of microbial kinetics under real reservoir temperature, pH, and salinity. That’s a feature, not a limitation; this is exactly the kind of physics you want grounding a safety-relevant forecast.
Where AI realistically enters is in making that expensive simulation affordable to run often enough to matter. A defensible souring study needs dozens of realizations — sweeping VFA content, abstraction capacity, injection chemistry, and treatment dosing — to produce the kind of probabilistic range the Field S/T numbers above show you actually need, not one more deterministic case. That is precisely the problem machine-learning surrogate models were built to solve across reservoir simulation generally: proxy models trained on a design of experiments through the full physics model, then used to screen thousands of scenario combinations in the time one full run used to take. Industry estimates for these surrogate techniques, reported broadly across history matching and forecasting workflows, run from roughly 10× to as much as 1,000× faster than the full simulation, with 10× to 100× more realizations fitting inside the same compute budget — figures that should be read as indicative rather than independently verified, but that match the general shape of what proxy modeling buys you anywhere it’s applied.
Applied to souring specifically, the realistic near-term role isn’t replacing a Field S/T-style compositional model with a black box. It’s training a fast proxy on an ensemble of runs through that model, so a souring risk range can be screened across every injector-producer pair in a field — and every candidate nitrate dosing schedule — in roughly the time it now takes to run a single full case by hand.
What to Actually Do About It
- Characterize background chemistry before injection starts — background H2S source (thermal vs. microbial), sulfate and VFA content, and the rock’s abstraction capacity, not after breakthrough already shows a trend.
- Budget nitrate dosing against the reaction zone, not the total mass balance — re-simulate dosing scenarios rather than scaling a dose linearly, since under-dosing can relocate the problem closer to your wells even as it shrinks on paper.
- Evaluate produced-water reinjection early, while there’s still time to build the transfer infrastructure it needs, rather than discovering after the fact that it was the most effective option you couldn’t use.
- Track H2S against water cut, not calendar time since injection start — souring tracks breakthrough timing at each producer, so the warning that actually matters is a forward prediction tied to each well’s own water-cut trajectory, not a field-wide schedule.
Common Pitfalls in Souring Risk Assessment
- Treating a single deterministic forecast as the answer. As the Field S/T base-versus-worst-case spread shows, nutrient availability and abstraction capacity are genuinely uncertain inputs; a risk assessment that reports one number instead of a range hides the question that actually matters.
- Assuming nitrate dosing is purely additive. The DTU zone-shift finding means a partial treatment can move the problem closer to the wellhead even while it lowers reservoir-wide totals — dosing has to be re-simulated at the rate actually planned, not scaled linearly from a bench test.
- Reading surface H2S as the leading indicator. By the time it shows up at the separator, the microbial community has been established for months to years; MPN, qPCR, and ORP trends in produced water move first.
- Ignoring produced-water reinjection because the infrastructure doesn’t exist yet. It was the most robust prevention option in the Field S/T study precisely because it avoids introducing fresh sulfate and nutrients in the first place — worth budgeting for before the next waterflood expansion, not after souring is already measured.
If you’re already managing the water side of a waterflood, our look at AI-based saltwater disposal injectivity decline prediction and AI waterflood conformance and thief-zone detection cover the two problems that most often show up alongside souring on the same injection wells.
