When the plant learns to run itself: reinforcement learning agents in desalination digital twins
Content by: Cristina Novo and Infographics by: Esther Martín Muñoz, Smart Water Magazine, May 13, 2026
When ACCIONA and Siemens commissioned the Al Khobar 1 desalination plant in Saudi Arabia in 2020, they did something the industry had never seen: they ran the entire start-up sequence remotely, from Madrid, through a digital twin built on the Siemens SIMIT simulation platform, replicating the entire hydraulic system from seawater intake to product water tank. With engineers unable to travel during COVID-19, the simulation platform held the plant together, virtually testing control programmes, running start-up sequences, and training operators before a single litre of water was produced. The plant hit its first-water milestone on schedule and went on to achieve energy consumption below 4 kWh/m³.
On the other side of the world, IDE Technologies' digital twin at the Carlsbad desalination plant in California, the Western Hemisphere's largest, uses five years of operational data to model membrane biofouling at the individual element level, projecting up to $1.5 million in maintenance savings over five years.
In Singapore, Gradiant's SmartOps AI platform is targeting a demonstration facility with energy consumption below 2 kWh/m³, against an industry benchmark of 3.5 kWh/m³.
These are the most sophisticated digital twin deployments in desalination today. None of them uses reinforcement learning (RL). That is about to change.
What digital twins do and cannot do
The current generation of desalination digital twins operates across three broad modes.
The first is engineering simulation: a physics-based replica of the plant used for commissioning, operator training, and control system testing, exemplified by ACCIONA's Siemens SIMIT platform at Al Khobar 1, which models every hydraulic system from seawater intake to product water tank.
The second is predictive monitoring: machine learning models trained on operational data to forecast equipment behaviour, detect fouling onset, and flag anomalies before they become failures. IDE's membrane degradation model at Carlsbad sits here, as does Gradiant's membrane cleaning prediction at PUB Singapore's Bedok NEWater Factory, a water reuse facility where the AI achieved 98.1% accuracy in predicting when cleaning was needed. In desalination specifically, Gradiant is working with PUB on a separate demonstration facility at Ulu Pandan targeting energy consumption below 2 kWh/m³.
The third is optimisation: AI tools that recommend or automatically adjust operating setpoints to reduce energy or chemical consumption. ACCIONA's ACRRO® system applies differential equation modelling to identify optimal configurations for RO racks, paired with a real-time ML tool called Insight that adjusts process parameters continuously, together forming what the company calls a dual-model optimisation system, deployed at a facility in Qatar. Gradiant's algorithms, derived from its 2022 acquisition of Canadian ML startup Synauta, deliver setpoint recommendations to operators at plants run by ENGIE, PUB Singapore, and others; at a large SWRO facility in the Middle East, they achieved up to 5% energy savings verified through the ISO IPMVP® measurement protocol, with savings reaching 18% at smaller plants in Australia.
Feature Image: Jason Jarrach on Unsplash