Can multiscale simulations paired with machine learning spell the end for trial and error and usher in a new era of data-driven research and development?

Serendipity is a word all of us have come across more than once when learning about the history of science. You’ve probably heard of Alexander Fleming discovering penicillin in a serendipitous moment after coming back from a holiday and finding mold growing on a Petri dish where he was growing bacteria.  

For centuries, science has heavily relied on serendipity to take big leaps forward. Some breakthroughs that were only possible thanks to the intervention of luck include the X-ray machine, the first chemotherapy drug and the discovery of graphene, among countless others.  

Outside of these ‘eureka’ moments, science mostly relied on trial and error. Constant, meticulous, repetitive, laborious experiments would be performed to obtain results that would then inform the next round of experiments.  

This approach, while effective at slowly chipping away at the barriers of our knowledge, had a big limitation. Without luck, trial and error cannot expand the bounds of human knowledge beyond just small improvements in areas we are already familiar with.  

Pushing the frontiers of our scientific understanding requires thinking out of the box and finding inspiration in unexpected places.  

The age of machine learning

Artificial intelligence has seeped into our day to day lives and integrated seamlessly into our smartphones and computers. This rapid takeover is rapidly and drastically changing industries across the globe, and is quickly making a dent in the world of scientific research.

With its unsurpassable ability to find patterns in vast amounts of data, machine learning is already making splashes in the world of scientific research.  

One of the main advantages of machine learning is that it can significantly speed up the trial-and-error process by sifting through millions of possibilities in record time. This new era of data-driven science is massively accelerating the rate at which we can conduct research.

But can AI be useful beyond trial and error and make serendipity obsolete?

Machine learning is now proving that it can help us go beyond our current knowledge. For example, AI has recently been used to lead the largest antibiotic discovery effort ever, finding close to a million new molecules that have potential antibacterial activity. Doing this through experiments alone would have taken many, many years.

This AI approach has already been used to identify a brand new class of antibiotic molecules — which has been a major challenge for the past 50+ years in the fight against antimicrobial resistance. Here, AI has enabled a departure from traditionally slow research workflows and brought forth a new era of data-driven insights.

The invaluable insights that AI can bring to any scientific field are starting to be recognized. Machine learning algorithms are now being tested in all kinds of scientific fields, anywhere from physics and biology to chemistry and materials research.  

However, like any other technology, AI has its limitations.  

For example, AI has been used to find the crystal structures of millions of materials. These crystal structures are essential to understanding the properties of a material, which then allows scientists to identify and design materials with optimal performance.  

Here, the question remains: are these structures feasible, or even useful?  

The key piece of information missing here is known as scientific understanding. Science is not just about the results of experiments — these are often the starting point to make predictions and formulate theories that explain how the world works.

In other words, AI can tell us what the correct result is, but not why it’s correct.  

At the end of the day, artificial intelligence is still a very new technology in the grand scheme of things. The way AI looks like today and how we interact with it will be completely different from what we’ll see in just a few years’ time.  

Computational simulations can help bridge the gap in understanding created by AI. Take the computational microscope — a technique that draws on the principles of atomistic simulations to create three-dimensional models of the behavior of every atom in a molecular system. This ever-evolving technology has been used in a variety of research areas, from biosystems to construction materials.  

For instance, it was recently applied to simulate the virus behind the Covid-19 pandemic and visualize its inner workings. This information can prove invaluable when it comes to developing vaccines and drugs against the virus.    

The key advantage of this approach is that it brings experimental data and AI-driven methods together to fill in the gaps in our understanding and create a comprehensive understanding of molecular systems.  

To really go beyond trial and error, we need traditional experimental approaches combined with a data-driven enhancement. Machine learning can help find an answer faster, while simulations can be used to validate the results and provide detailed information on how to proceed in the lab. This process can be repeated in a loop to keep advancing our knowledge at a much faster pace than we’ve ever seen before.  

Machine learning has already shown its potential to surpass the role that serendipity has had on the progress of science over the course of history. Still, it might not yet be the end for serendipity.

The future of serendipity in science

While machine learning can offer scientists a shortcut through processes that in the past would have required a stroke of luck, serendipity is unlikely to become obsolete.  

Rather, serendipity — and the role it plays in science — could get a major boost from AI technologies.  

The truth is that serendipity in science was never just about luck. It’s always been a mix of creativity and luck that distinguishes real innovation from mere iteration.  

When Fleming discovered penicillin, luck was definitely involved. But it was his curiosity that led him to study the mold he found on that petri dish, instead of just throwing it away without a second thought.  

Andre Geim, one of the discoverers of graphene, is known for saving time every Friday to perform playful experiments outside of his main research focus. Graphene was the result of one of these experiments, where Geim and his colleague Kostya Novoselov were playing with pencils and Scotch tape.  

Curiosity and a creative mindset, sprinkled with some luck, is what has propelled some of the biggest scientific revolutions of the past.  

Going forward, AI and serendipity don’t need to work against each other, but together, to propel us forward faster than ever before.  

AI could be a source of not just data, but also of inspiration for new concepts. Humans can then take these ideas and develop them further into theories to expand our scientific understanding.

For example, the inhuman ability of AI to find patterns in data can be exploited to find unexpected results within large datasets, or even within the entirety of the scientific literature published on a specific topic.  

Scientists are already developing and testing AI models focused on extracting new concepts from the available data and providing explanations for the results — this is a rapidly growing branch of artificial intelligence research known as explainable AI, or XAI.  

In the age of AI, serendipity can evolve to adopt a new, amplified role in our search for the unknown, working in close collaboration with multiscale simulations and machine learning technology to move beyond trial and error.

‍