Manchester researchers harness physics‑informed AI to accelerate carbon removal assessments

Researchers at the University of Manchester are turning to physics‑informed artificial intelligence to accelerate assessment of carbon dioxide removal (CDR) strategies, seeking to generate evidence faster and at lower cost than traditional field trials and numerical models. According to the report by myScience, the team’s work aims to model the global carbon cycle and virtually test interventions such as ocean fertilisation, ocean alkalinity enhancement and enhanced rock weathering before committing to large‑scale experiments.

Field experiments for nascent CDR approaches are slow, expensive and carry environmental uncertainties. “Field experiments, especially for ideas such as ocean fertilisation, are costly and slow. With the advent of physics‑informed AI, we can replace or facilitate such experimental campaigns with predictive models that can incorporate a more accurate representation of physical processes than common numerical models and are also faster. This enables us to study proposed CO2 removal methods at scale,” Dr Peyman Babakhani, lecturer in Geo‑environmental Engineering, told myScience. The University of Manchester similarly described the initiative as using physics‑informed AI to provide faster, more flexible predictions than classical numerical approaches.

The Manchester work blends nanotechnology research with AI: one strand explores engineered nanoparticles, iron, silica or aluminium, as a way to boost phytoplankton growth and increase carbon export in ocean fertilisation scenarios. Laboratory and in silico evaluation, the team argues, are essential before any field deployment because nanoparticles may alter bloom dynamics, ecological interactions and biogeochemical cycles in ways that carry environmental risks.

AI methods being applied are not limited to conventional neural networks. Recent advances in physics‑informed neural networks and hybrid models point to major efficiency gains: a development paper on AQ‑PINNs, an attention‑enhanced quantum physics‑informed neural network, reports a 51.51% reduction in model parameters while retaining comparable convergence and loss for fluid‑dynamics problems governed by the Navier‑Stokes equations. According to the arXiv preprint, that architecture also reduces computational burden and the model’s carbon footprint, attributes of direct relevance when modelling Earth‑system processes at scale.

The potential industrial value is twofold for decarbonisation stakeholders. First, faster, uncertainty‑aware virtual assessments can triage which CDR concepts merit costly pilot programmes. Industry data and modelling consortia frequently cite time and capital as major barriers to translating promising lab results into deployable solutions; AI‑accelerated evaluation could reduce both. Second, by embedding physical constraints into AI, models can produce more defensible policy and investment guidance than black‑box approaches, helping operators, investors and regulators weigh trade‑offs between efficacy, cost and environmental risk.

Other groups are pursuing complementary physics‑aware AI for climate mitigation infrastructure. A Penn State team has combined physics and deep learning to assess subsurface CO2 storage suitability more efficiently, according to a university release; such techniques can streamline identification of secure sequestration sites and lower technical barriers to carbon capture and storage (CCS) deployment. Manchester researchers have also applied AI to related climate domains, discovering new wave‑breaking equations and building city‑scale digital twins to plan equitable net‑zero transitions, illustrating a broader institutional push to couple domain science with machine learning.

Despite the promise, the researchers and surrounding literature emphasise caution. Modelling can never fully substitute for empirical validation where ecological complexity and long‑term impacts are concerned. The Manchester reporting stresses that AI models are intended to “replace or facilitate” experimental campaigns, not to circumvent rigorous laboratory and phased field testing. Where competing accounts or uncertainties exist, over nanoparticle behaviour, carbon export efficiency or unintended ecosystem effects, those differences should be resolved through coordinated laboratory work, targeted mesocosm studies and tightly regulated pilot deployments informed by model outputs.

For industrial decarbonisation practitioners, the Manchester approach signals a maturing toolkit: physics‑informed AI can lower the cost and time to screen CDR concepts, improve scenario analysis for investment decisions and produce uncertainty estimates that are actionable for risk management. Yet successful uptake will require transparent model documentation, independent verification, and integration with monitoring frameworks to ensure that virtual predictions align with observed outcomes as projects scale.

Dr Babakhani’s team, a member of the ExOIS (Exploring Ocean Iron Solutions) Forum, positions these AI tools as an early‑stage filter, helping industry and policymakers prioritise interventions that merit the expense and scrutiny of real‑world trials while highlighting those with unacceptable environmental risk. As the field evolves, combining improved physics‑aware algorithms, laboratory evidence and cautious field experimentation will be essential to move promising CDR technologies from concept to credible, scalable climate mitigation options.