Fluid simulations today are remarkably realistic. In this Comment I discuss some of the most striking results from the past 20 years of computer graphics research that made this happen.

Growing up, I would often marvel at the smoke plumes ascending from a chimney and the water flows in the wake of a ship and assume that the underlying rules that describe them must be unfathomably complex. Later, as a student, I was struck by the deceptive simplicity of the Navier-Stokes equations, which—using only 3 terms codifying advection, pressure and diffusion—could describe waterfalls, waves around water droplets and turbulent smoke swirls.

Understandably, computer scientists have been eager to plug these equations into a computer and see the world come to life in their simulations. Unlike the computational fluid dynamics1 literature, which aims for rigorous and accurate results, computer graphics research typically focuses on greater efficiency and artistic control, which are achievable with approximate solutions. These graphics solutions started appearing over 20 years ago^2,3, but hundreds of papers on this topic were still published in these 20 years—a testament to the complexity of the problem.

…Even with these improvements, it always seemed that in computer graphics, this realism is only there for looks—and I never expected these simulations to have any predictive power. But today, this prospect is becoming more and more likely. For decades, physics simulations for digital media were considered acceptable if they looked convincing to the human eye, and were nowhere near accurate enough for engineers to verify, for example, whether a new wind turbine design really does work correctly.

However, the computational cost of existing methods has decreased 4× in just one year due to simpler, more efficient geometric approximation schemes that map more easily to existing graphics cards. With this, one can now simulate the airflow within a city block or create predictive wind tunnel tests for aircraft wing design with each second of animation taking only a few minutes to compute⁹. Simulations that are both real-time and predictive are within arm’s reach—we might soon enter a world where an engineer is able to test new ideas in aircraft design every few minutes. There are many more techniques that enable the simulation of intricate fluid phenomena such as the mesmerizing phenomenon of a ferrofluid climbing up a steel helix¹⁰ or liquid-hair interactions¹¹. It is also possible to extract the physical properties of a viscous material from a video recording of its dynamics¹².

These technical advances come at a price. More complex systems don’t map well to existing hardware and their code base is more difficult to maintain and troubleshoot over time. Striking the right trade-off remains a key challenge when developing new simulation algorithms. However, this is also what makes this area a fertile ground for new ideas, where a small, but well-chosen compromise can introduce an order of magnitude increase in efficiency: those are the landmark papers in computer graphics.

On the other hand, neural-network-based learning approaches can generate increasingly convincing physics simulations more and more efficiently with each passing year¹³. Over time, they may even surpass conventional simulation methods. However, creating a unified system that can accommodate our appetite for realism and leave space for artistic directability, and do so efficiently enough to fulfil the requirements of modern artistic workflows, remains a challenge.

I would like to think that I have a vivid imagination, but after seeing all this progress I wonder what else we will be capable of 20 years and a few more papers down the line.

See Also: