AI and flexible investment model lifts simulated wastewater biorefinery value

A simulated wastewater biorefinery became economically viable in a modeling study that combined a recurrent neural network with real options analysis, an approach designed to preserve managers’ ability to delay or expand investment as conditions change. Over a 10-year simulation, the AI-enhanced framework produced an extended net present value of $2.8 million, compared with a negative $1.2 million under conventional static net present value analysis. The research, published in Frontiers in Environmental Engineering, addresses a central problem in wastewater valorization: recovering biogas, phosphorus, nitrogen and clean water from sewage can generate products with market value, but influent chemistry, commodity prices and operating conditions vary over time.

A conventional project appraisal generally commits the full investment at one point and discounts expected cash flows under fixed assumptions. That approach can reject projects that would become worthwhile if operators retain the right to wait for better information, expand when prices rise or scale output when resource availability improves. The study therefore compared a static appraisal with two models that explicitly represent flexible decisions.

To represent those choices, the authors developed an AI-Enhanced Real Options Theory framework, or A-ROT. Its forecasting layer used a Long Short-Term Memory network to predict a combined resource value index derived from daily biogas and struvite yields and their respective market prices. The model also included influent flow, electricity tariffs and fertilizer subsidies as inputs.

A real options layer then treated the biorefinery as an American call option: management could defer an irreversible capital commitment, exercise the option when conditions became favorable, or choose later expansion. Monte Carlo simulation supplied 10,000 possible market and operating paths, while explainable AI tools traced the influence of individual variables. The forecasting dataset contained 5,000 daily observations spanning January 2012 to September 2025.

Historical wastewater and market information was used to calibrate and train the models, whereas the forward-looking investment analysis was simulated over 10 years. The operating data were reconstructed synthetically from baseline biokinetic and flow values associated with a metropolitan treatment facility, while commodity indices were mapped to historical spot markets. The dataset was divided chronologically: 3,500 observations for training, 750 for hyperparameter tuning and validation, and 750 as a held-out test set.

This chronological partition was intended to prevent information from later periods leaking into earlier model training. Data preprocessing included localized linear interpolation for missing values, which represented less than 0.4% of the dataset, normalization, outlier detection and temporal alignment. Values exceeding three standard deviations from a rolling 30-day mean were winsorized to limit distortion during neural-network training.

The simulation separately assigned stochastic distributions to key variables. Phosphorus and natural-gas price indices followed log-normal processes, influent flow followed a mean-reverting Ornstein-Uhlenbeck process with a lower bound of 8,500 cubic metres per day, and effluent volatile suspended solids loading was represented as Gaussian with a mean of 220 milligrams per litre. The model’s forward horizon was a simulation rather than a historical investment record.

The LSTM architecture contained two hidden layers with 128 and 64 memory units, each followed by 20% dropout, and a dense output layer that produced 30-day forecasts. It was optimized with the Adam algorithm at an initial learning rate of 0.001, using mean squared error over as many as 150 epochs and early stopping with a patience of 15 epochs. It was compared with an ARIMA baseline, a random forest regressor and XGBoost using the same features and test sequence.

The LSTM achieved a mean absolute percentage error of 4.2%, an R-squared value of 0.94 and a root mean squared error index of 0.051. ARIMA’s MAPE was 14.8%, while random forest and XGBoost produced MAPEs of 8.1% and 6.85%, respectively. Feature analysis identified effluent volatile suspended solids as the most influential variable, followed by natural gas futures, influent phosphorus concentration, electricity tariffs and regional fertilizer subsidies.

Volatile suspended solids served as a proxy for organic load available for anaerobic conversion into biogas, while phosphorus concentration represented a physical constraint on struvite production. SHAP values and mean-decrease-in-impurity scores were used to interpret these rankings. The model therefore connected physical treatment conditions with commodity markets rather than treating resource recovery solely as an engineering problem.

In the valuation layer, forecast uncertainty was linked to a local volatility measure for the option model, allowing predictive performance and market uncertainty to change the value of waiting. Valuation changed substantially when managerial flexibility and predictive information were introduced. Static net present value and the basic real options model both returned a static project value of negative $1.2 million.

Basic real options added $1.85 million in flexibility value, lifting extended NPV to $0.65 million and moving the optimal exercise point to year 1.5. A-ROT added $4 million in option value relative to the static baseline, raising extended NPV to $2.8 million, moving the modeled exercise point to year 2.4 and increasing the internal rate of return from 6.8% to 18.4%. These differences arose within the model’s assumptions rather than from a controlled comparison of actual facilities.

The framework also examined the effect of better information on financing costs. A simulated conventional weighted average cost of capital of 14.5% fell to 9.8% when the AI component was included, although the authors explicitly described this as a scenario assumption rather than a guaranteed response by lenders or investors. In the 10,000-iteration Monte Carlo analysis, the probability of a positive NPV was 84% under A-ROT and 31% under traditional NPV.

Mean extended NPV was $2.80 million, with a reported 95% confidence interval of $2.45 million to $3.15 million, while 95% Value-at-Risk fell by 42%. The model also tested correlated commodity-price and wastewater fluctuations rather than varying each input independently. Those modeled gains depended on market conditions and forecast quality.

At 10% commodity-price volatility, the model assigned $1.1 million to strategic option value and favored early exercise; at 25% volatility, value rose to $2.8 million and favored waiting for an AI signal; at 40%, it reached $4.5 million and favored retaining the option. Forecast error was equally consequential. A model with 2% MAPE generated $3.2 million in extended NPV, the study’s 4.2% forecast generated $2.8 million, and a 10% threshold produced $0.4 million.

Beyond that threshold, the additional value attributed to AI diminished and the valuation moved toward the traditional result, establishing 10% MAPE as a modeled threshold for economic added value rather than a universal requirement. The simulations also exposed a weakness in data-dependent decision systems. The researchers introduced a simulated data-poisoning bias into sensor inputs for volatile suspended solids over 60 days.

The corrupted forecasts caused the option model to expand in year 1.1 rather than its preferred year 2.4, leaving infrastructure underused when actual wastewater volumes failed to support the added capacity. Extended NPV fell from $2.8 million to $2.38 million, a reported 15% contraction. The scenario shows that forecast quality and the integrity of operational data directly affect the modeled investment decision.

It does not establish that cybersecurity failures are as consequential as, or more consequential than, the underlying process technology. Overall, the study presents a decision-support framework rather than proof that AI can finance or operate real wastewater biorefineries. Its strongest finding is conditional: in the simulated case, pairing accurate 30-day forecasts with flexible investment decisions changed an apparently unviable project into one with positive modeled value.

The approach gives managers a way to connect the physical composition of wastewater, expected commodity returns, market volatility and the timing of capital spending. It also identifies predictive accuracy and trustworthy sensor inputs as necessary conditions for that modeled economic benefit. Because the biorefinery, historical calibration series and forward market paths were partly or wholly simulated, the results demonstrate the behavior of the proposed framework rather than verified commercial performance in a municipal treatment plant.

Subject of Research: AI-enhanced real options modeling for investment decisions in simulated wastewater valorization projects

Article Title: The economics of circularity in wastewater valorization: harnessing ai and dynamic modeling for optimal investment, pricing, and resource allocation

Article References: Yusuf, J. A., & Adejoh, S. (2026). The economics of circularity in wastewater valorization: harnessing ai and dynamic modeling for optimal investment, pricing, and resource allocation. Frontiers in Environmental Engineering, 5, 1822946. https://doi.org/10.3389/fenve.2026.1822946

Image Credits: AI Generated

DOI: 10.3389/fenve.2026.1822946

Keywords: wastewater valorization, real options theory, LSTM forecasting, biorefinery economics, dynamic valuation, Monte Carlo simulation, investment timing, circular economy, data poisoning, explainable AI, economics, circularity