Artificial intelligence is taking a leap forward with new tools that take us well beyond writing essays and summarizing documents. While large language models (LLMs) are adept at text and visual learning models (VLMs) are excellent at producing graphics and short videos, another set of AI tools is needed to address the most pressing challenges of our time: developing medicines for diseases, accelerating the move to renewable energy and storage, creating new materials for automotive and aerospace, developing safer and more efficient water filtration and tackling other key goals.
That’s where AI simulation steps in. AI simulation combines techniques from quantum physics with the power of deep learning to address these challenges.
Let’s first focus on biopharma development. McKinsey published research showing that “the average time for taking a new medication from candidate nomination to launch has been about 12 years.” As for cost, a Deloitte report found that the “average cost of developing a new drug rose by $298 million to $2.3 billion in 2022.” We need cost-efficient and timely drug development that produces safe medicines. The sooner we have them, the faster we can positively impact patients' lives.
Other pressing issues we must solve include bringing new battery chemistry to power EVs and storing renewable energy at scale. We also need a way to remove forever chemicals (PFAS) from water filtration and other key areas of our lives. We have to go beyond just text or image-based AI to crack these larger problems.
Artificial Intelligence models such as ChatGPT do well when they are trained on large amounts of data. What do we do, though, when there isn’t a lot of data—which is the reality in the majority of cases in the physical world? For instance, if we had lots of data on a new drug for Alzheimer’s disease, we would have already solved the problem. What is needed is a new form of AI that can generate high-quality data based on real-world physics and dynamics. This is where simulation comes into the picture. A simulator is a computer program that contains the key equations that drive molecular or other dynamics and can run the possible combinations and interactions billions of times. Deep learning AI is then applied to this data to optimize for a desired outcome, such as a drug with high binding affinity to a target receptor.
Running billions of simulations of a digital twin of the molecule in a large GPU cluster computer model is far easier, faster and less costly than manipulating a physical experiment in a lab. Along the way, AI models can learn from past simulations to further improve.
Consider ConPLex, a model created by MIT and Tufts University researchers that, according to MIT News, enables them to “screen more than 100 million compounds in a single day.” Simulation is particularly helpful for advancing breakthroughs in challenging, undruggable conditions like cancer or Alzheimer’s, where little data exists for AI to model. In this case, the quantum mechanical interactions between drug compounds and human receptors—simulated on today’s classical computing hardware (GPUs)—create new data from physics and unlock insights that have eluded researchers for decades.
Using simulation, researchers can create drugs that are safer for patients and have a better chance of succeeding at clinical trials by running virtual scenarios of drug interactions within the human body before the first person takes the drug. And they can do it faster and much more cost-effectively by eliminating years of painstaking lab work trying to replicate molecular-level interactions to find the most promising compounds. This ability to impact the physical world is why some GPU manufacturers are doubling down on these new forms of AI.
The combination of AI and simulation goes beyond medicine. Combining the two technologies helps scientists and others make cutting-edge discoveries in various areas with less time, money and risk than traditional avenues.
Clean energy is another leading use case. While lithium-ion batteries have been integral in powering electric vehicles and helping to store solar and wind energy, they have major limitations. We need batteries that have higher energy density, weigh less and are less expensive to produce. However, performance-testing advanced battery chemistries, materials and designs can take years via traditional methods. Using simulation, researchers can explore countless combinations by using AI to optimize designs and make performance predictions much faster, helping to speed these new batteries to market or optimize them for specific applications.
Other key use cases include materials science, food technology and cosmetics product development. If it’s a tangible item, chances are that AI and simulation can work together to help researchers create a better version of it. For example, using these two technologies, scientists can engineer new types of biodegradable plastics or more sustainable construction materials, minimize harmful additives in the foods we consume, and create skincare and beauty products that do not have some of the toxicities we see today.
AI simulation opens up many new possibilities to tackle large problems. To validate the results of these computer models, we must still put the final suggested molecular structures through real-world testing. There is a continual feedback loop between AI simulation and physical testing that drives progress forward. This is critical not only for safety but also for improving the AI models in use; the more reinforcement learning that researchers program into the development process, the quicker the models will improve.
By strategically using AI and simulation together now, we can make breakthroughs that will create commercial and societal benefits. Key to this success will be the training of engineers and leaders in these novel AI tools. University, corporate and governmental leaders all have roles to play in driving this upskilling. Let’s make sure these powerful tools are democratized across the world so all can benefit from them.
