Editor’s note: This commentary was adapted from Dirk Hartmann’s discussion, “Living in a Simulation,” on the Human & AI podcast.
We live in a time when machine learning and simulation are both getting many times more sophisticated, accurate, and fast. The question the two technologies pose is obvious:
How can we combine machine learning and simulation to bring them closer together in way that follows what Isaac Newton was doing in his home office in the years 1665 and 1666, when Cambridge University was closed due to the Great Plague?
And what was Sir Isaac getting up to in his laboratory? He was developing his first ideas towards his breakthroughs of calculus, optics, and the law of gravitation, which have led to a deeper understanding of physics.
This kind of extended, combined thinking is exactly what is driving me as a researcher: How well can we really develop a deep understanding of and predict the behavior, or the outcome, of physical systems and use this to optimize both products and operations?
In the recent past, few people were really trying to combine these two technologies. Some technicians were using simulation models to collect additional data for a machine-learning model. But in the last couple years, with the capability of algorithms having grown exponentially over the past decade, researchers like me have seen the rise of something called physics-informed neural networks, the melding of machine leaning and simulation.
Essentially, researchers are bringing the physics information into the neural networks architectures and their training functions to create this combination of neural networks and physics models. What this means practically is that many real-world applications that cannot be realized due too little and clean available data become feasible through simulated physical insights.
Developing physics-informed neural networks is quickly becoming a vast new field with a lot to teach us, because we can go so far beyond classical applications for simulation, mostly in design and engineering, and open up new paradigms for how we operate our equipment. Specifically, we can run a simulation in parallel to the actual operation of an asset, such as an electric motor. We can use the simulation to infer internal values that could not be collected using sensors. That brings a much more accurate understanding of performance values that were previously unknown, and that enables completely new operations strategies.
With such information, we could teach a robot how to do milling tasks that previously were not possible. A simulation running parallel to the operation of the robot, combined with data optimization, could increase accuracies to 0.1 millimeter. That's an increase at industrial scale; i.e., such an increase in accuracy means that robots can now be used for industrial milling tasks, replacing large gantry machine tools.
We can run a simulation in parallel to the actual operation of an asset, such as an electric motor...[using] the simulation to infer internal values that could not be collected using sensors.
Even more practical, given our current situation, physics-informed neural networks can help reduce the impact of COVID-19. Researchers can perform computational fluid-dynamics simulations to model how people coughing leads to a spread of the virus. This can answer questions about the use of face masks and the value of social distancing, and how to optimize these tactics.
Such pandemic-related modeling also has significant use for industry. We can use these computations to create simulations that predict how people move in industrial and manufacturing spaces. This can lead to optimizing the path people follow on the shop floor so that they automatically increase and maintain their social distancing.
Beyond that, there are potential medical applications. Can we simulate the air flow within human lungs to model how the SARS-CoV-2 virus affects the lungs and plot our way to an effective treatment? Can we then use mass amounts of data, from both real-world actions and simulations, to simulate how a cure should work? These are major questions on a lot of researchers’ minds.
Combining the best of physics-based modelling and simulation on the one hand and machine learning on the other hand, we are today able to gain insights into systems of a complexity not imaginable a few years ago. As our technical capabilities increase, we will soon be able to automatically formulate our understanding of beautifully complex physics much the way as Newton did.
Published on August 20, 2020