AI grounded in physics is cracking fluid dynamics’ toughest puzzles, revealing new singularities and reshaping how science and industry solve complex problems.

Executive Summary 

For centuries, the equations of fluid dynamics have both illuminated and frustrated humanity’s understanding of the physical world. They are among science’s most powerful tools, underpinning natural phenomena from aerodynamics to the movement of subterranean hydrocarbons. Yet they are riddled with singularities, points where the formula collapses into infinite values. A new approach, pioneered by Google DeepMind that uses Physics-Informed Neural Networks (PINNs), has uncovered entirely new families of these singularities, pushing the frontier of what’s possible in mathematics, physics, and engineering. This marks the arrival of physics-grounded AI as a serious tool for discovery, bridging theory and practice in ways traditional methods could not. 

The Century-Old Puzzle of Fluid Dynamics

Fluid dynamics has long been one of science’s great enigmas. These equations describe how air swirls over a jet’s wing, how water churns in a storm surge, or how oil and gas move through porous rock. 

Yet hidden within these mathematical models are paradoxes: points where pressure or velocity explode into infinity. These singularities, often called ‘blow-ups’, don’t occur in reality. They do, however, expose the limits of these models. Studying them helps sharpen the boundary between what mathematics predicts and what physics actually permits. 

One such puzzle, the Navier–Stokes equations, is so fundamental that proving whether singularities exist within them is one of the Clay Millennium Prize Problems, with $1 million on the line. 

Why Unstable Singularities Matter

Change management in the AI era cannot be reduced to workflows and dashboards. It must account for the human variable. Workforce shifts ripple outward, influencing supply chains, local economies, and entire professions. Entry-level programmers encounter code-generating models, multilingual workers engage with real-time translation tools, and educators adapt to AI platforms that personalise lessons more effectively than static textbooks. Beyond knowledge work, sectors like transport, logistics, and farming are preparing for similar transformations. 

AI in healthcare offers a valuable lesson. Most successful applications of AI in this sector have not come from replacing doctors, but from automating routine tasks so that scarce experts can focus on higher-value decisions. The same principle applies to every industry: productivity gains must enhance human judgement, not sideline it. 

The Question of Free Choice

Not all singularities behave the same. Some are stable, resistant to small changes. Others collapse under the slightest shift. The new DeepMind research mapped these unstable singularities systematically for the first time. 

This matters because in the Navier–Stokes equations, most experts believe stable singularities may not exist at all. That leaves unstable ones as the only pathway forward. By uncovering new families of them, researchers have opened the door to questions that have resisted centuries of effort. 

A Physics-Grounded Approach to AI

The breakthrough lies in how AI was applied. Most systems today, whether large language models or computer vision tools, are trained on oceans of data. They excel at spotting patterns but falter when reality strays from their training examples. 

Physics-Informed Neural Networks (PINNs) invert this approach. Instead of learning from data alone, they are trained against equations of physics. Every prediction is checked against the governing laws of motion, and the network learns by reducing that mismatch. 

This is the essence of physics-grounded AI: intelligence that is tethered not to correlations, but to the laws of nature. 

The parallels with energy operations are clear. General AI fails in the messy, high-stakes environments of refining. Only specialist, physics-informed systems can operate safely and effectively at scale. The same lesson holds here. AI rooted in first principles unlocks capabilities, such as preventing hallucinations, where brute-force pattern-matching struggles. 

Precision at the Edge of Discovery

DeepMind’s team didn’t just use PINNs to formulate new frameworks; they pushed these models to near-machine precision, akin to measuring Earth’s diameter within a few centimetres. This accuracy enables these neural networks to contribute to rigorous, computer-assisted proofs. 

The result is the systematic discovery of unstable singularities across multiple fluid equations. Even more intriguingly, patterns emerged linking the speed of singularity blow-up to their instability order, hinting at deeper structures still to be uncovered. 

This isn’t AI accelerating research. It’s AI expanding the frontier of mathematics itself. 

From Fluid Equations to Industrial Systems

At first glance, singularities may appear as academic jargon. But they embody what happens when theory breaks under real-world conditions, the same challenges operators face in complex systems like refineries. 

Today, most refineries use less than 10% of their available data for decision-making, leaving vast value untapped. By contrast, Orbital uses 100% of operational data, harnessing time series, physics-based, and language models to generate real-time insights. The opportunity is staggering: tens of millions in hidden value, waiting to be unlocked. 

The bottom line is evident. Aligning AI with physical laws isn’t just academic; applied AI, grounded in the realities of chemistry and physics, delivers tangible performance and profitability gains by unlocking the next frontier of downstream optimisation. 

Orbital and the New Era of Physics-Based Models

This very philosophy underpins Orbital, the foundation model built for energy operations. It fuses three capabilities: time series analysis, physics-based modelling, and decades of chemical and process engineering expertise embedded in a language model. 

The result is a system that reasons in real time, grounding insights in physics as well as data. The same principles that revealed hidden singularities in fluid equations at DeepMind now make possible optimisation breakthroughs in industrial AI. 

Orbital uses 100% of operational data, harnessing time series, physics-based, and language models to generate real-time insights that unlock tens of millions in hidden value

A New Era of Fluid Dynamics

Fluid dynamics was once the preserve of mathematicians with chalkboards. Today, with physics-grounded AI, it is a proving ground for a new form of intelligence, one that respects the laws of nature while extending human discovery. 

The implications ripple far beyond equations. From hurricanes to pipelines and jet engines to refineries, physics-grounded intelligence shows that the path forward lies not in abstract generalisation, but in specialisation, precision, and adherence to laws of the physical world.