According to all the available evidence, new ideas are getting harder to find.
Since the 1990s, the growth in novel patents — those that mention a new technology — has been negative in the United States. According to one widely cited study, U.S. research productivity declines 50% every 13 years, largely because new ideas are drying up. The implication is that the U.S. needs to double its research investment about every dozen years just to stand still in growth terms.
But the growing difficulty of discovering new ideas — the next blockbuster drug for cancer, the next graphene — may come as no surprise when we consider that the odds of success are heavily stacked against U.S. scientists in the first place. Every material on Earth is made up of some unique combination of the 118 elements in the periodic table. Trillions of combinations remain undiscovered. Of these, most have no useful properties for industry. Finding the tiny subset of useful new materials is like trying to find a needle in a field of haystacks.
Even when promising candidates are identified, the vast majority fall at the later hurdles of testing, regulatory approval, and development. In health care, for example, only 1% to 2% of promising drugs make it to the market, with costs in the billions of dollars spiraling year on year.
New technologies such as machine learning, robotics, digital twins, and supercomputing can dramatically improve the odds of successful discovery — through faster and more efficient selection of the most promising new ideas and accelerated testing to get these ideas into development sooner. These technologies are paving the way to breakthrough advancements in fields as diverse as energy, medicine, and urban planning.
Machine-Powered Discovery of New Materials
“We’re used to thinking about machines as doing the jobs that are dull, dirty, and dangerous, whereas humans do the creative stuff,” technology writer Luke Dormehl, author of Thinking Machines, told us in an interview. But this line between humans as creators and machines as executors is blurring: “In practice, a lot of human creativity comes from previous ideas or brute-force experimentation — a musician might try scores of chords before finding a new tune, for example.” Moreover, “machines can often suggest ideas or solutions that humans wouldn’t think of, suggesting that they are capable of creativity in the broad sense.”
This is perhaps clearest in the search for new chemicals and materials, where discovery has traditionally been hypothesis driven: Scientists make informed guesses about which chemical compounds or materials might work, which then must be tested and retested extensively. The approach has many drawbacks: high rates of false positives, waste of expensive materials consumed in testing, and long wait times for results.
Kebotix, a Cambridge, Massachusetts-based company, disrupts this scientific method. Dubbed the world’s first self-driving lab for materials discovery, the company uses a combination of computational modeling, robotics, lab automation, and machine learning to accelerate the rate of discovery of advanced chemicals and materials.
Jill Becker, a Harvard-trained chemist and the CEO of Kebotix, explains: “We aim to transform the centuries-old, manual scientific method. Our AI can rapidly and efficiently process enormous amounts of complex molecular or chemical data to discover new materials or generate new formulations of particular products with desired target properties — dramatically condensing the research cycle from years to months.” Consider the example of smart windows for a new car model: “We can specify that the material has to be transparent, that it has to let light in, but not heat, in summer, that it has to save on energy and environmental costs, and so forth.”
Becker sees the bigger picture, too: “These technologies can help address some of the world’s biggest problems, for example by speeding up the search for cleaner, greener materials or finding new health treatments.”
New Health Care Treatments
Some of the most exciting developments are happening in the pharmaceuticals sector, where the search for new blockbuster drugs using traditional methods has become ever harder and more costly. British pharma-technology company Exscientia tasked its AI platform with designing a drug to treat obsessive-compulsive disorder. After sifting through millions of potential chemical structures, Exscientia decided to make and test 350 compounds, only one-fifth of the usual number of candidates. In January 2020, an AI-designed drug reached the clinical trials stage after just 12 months — a very rapid time frame compared with the usual four and a half years.
Machine discovery is not just used to find completely new treatments; it often provides faster canvassing of existing therapies to find better solutions to new diseases or medical conditions. A case in point is the scramble for fast and effective treatments and vaccines during the COVID-19 pandemic. U.K.-based biotechnology company Benevolent AI used its algorithms to search a database of existing drug therapies, ultimately identifying a successful candidate in the form of a drug currently used for rheumatoid arthritis. The treatment has few side effects and works in combination with other therapies.
A major challenge for machine discovery is the enormous computing power required for large-scale searching. One solution to this constraint is the VirtualFlow drug-discovery system developed by a cross-disciplinary team of scientists. As described in a recent Nature article, the open-source platform aims to tackle the problem of scale in finding useful drug candidates. While existing databases reference over 1.4 billion commercially available compounds, previous methods could search only a small subset of these compounds due to the immense computing power required. Using elements of parallel computing and cloud systems, the VirtualFlow platform runs parallel searches across the chemical space — with staggering implications. Using a single computer processor, it would take approximately 475 years to search 1 billion chemical compounds; the VirtualFlow system harnesses the power of over 10,000 linked computers to search the same number in two weeks.
Machine discovery is also helping in the fight against climate change by accelerating the search for cleaner energy sources. Researchers at the University of Liverpool used an AI-guided robot to identify stronger catalysts for the production of hydrogen from water, which can then be used in hydrogen fuel cells that have a much lower environmental impact than fossil fuels. Roaming freely around the laboratory while manipulating an array of different instruments and vials, the robot was able to perform 688 experiments over eight days — 1,000 times faster than traditional manual laboratory methods.
Curtis Berlinguette, professor of chemistry and chemical and biological engineering at the University of British Columbia, cautioned that machine-enabled discovery won’t apply to every scientific problem: “So far, it has worked better for small-molecule chemistry, such as drugs, than for materials.” Berlinguette is interested in applying machine learning and automation to thin-film technologies. “Every clean energy technology relies on thin films — solar cells, batteries, smart windows, even electrolyzers — to turn CO2 into fuels.” Berlinguette and his team designed and built a robot tasked with creating defect-free films for solar panels; this includes the analysis and testing of various candidates to determine the best material to use. In just five days, Berlinguette’s robot can complete a process that once took nine months.
Outside the lab or the design studio, however, real-time ideation and testing become harder. Consider the problem of designing better cities: Because of the difficulty of controlling for all variables, such as environmental conditions, it’s hard to test ideas like a new flood control or traffic routing system.
Yet even here, help is at hand in the form of the digital twin, a virtual replica of an object, being, or system that can be continuously updated with data from its physical counterpart. Phil Christensen, vice president in the Digital Cities department at Exton, Pennsylvania-based software development company Bentley Systems, told us, “Digital twins can serve as a sandbox environment for large-scale simulations. Take environmental change: We can simulate what the impact of different rainfall patterns would be and how floods would impact a city. We can simulate new traffic plans and examine the changes in traffic flow and pollution. Digital twins are set to become invaluable assets for experimentation and decision-making.”
Getting Ready for Machine Discovery
Despite these advances, capitalizing on machine-powered discovery and testing may be harder in practice for a number of reasons. Scientists may feel threatened by the changes; the up-front investment is significant; and commercial models may be out of sync with machine discovery. Three actions can help organizations and industries successfully harness the power of machine-powered discovery.
Make ideation a human-and-machine team sport. It would be a mistake to see machine discovery as simply automating some tasks, with humans supplying the creative spark. The model should be more like an innovation tag-team, with humans and machines interacting continuously according to each other’s cues and insights. To be sure, as technology writer Dormehl observed, humans will probably guide the initial research questions, but machines can complement and improve these along the way, often by finding patterns that humans wouldn’t detect. Berlinguette noted that his team uses machine learning algorithms that start by searching combinations of ingredients and processing steps suggested by the researchers, but also identify less obvious combinations that human scientists may overlook.
Kebotix’s Becker underlines the team sport aspect of machine discovery: “I see a future where scientists can dream more and be more curious, bantering back and forth with the AI about ideas.”
Human and machines think in different ways, but it’s the collaboration between them that produces results. Implementing machine discovery requires a genuinely cross-disciplinary approach involving not just physicists or chemists, but technologists, data scientists, software engineers, ethicists, and product developers.
Find the technology backbone. Machine discovery depends on a complex array of different technologies — machine learning algorithms, sensors and smart objects, robotics, reference databases, and high-speed computing networks — all in constant communication with one another. Not every business or lab will have the funds to invest in their own in-house systems; to supply the necessary technology backbone, smaller organizations can consider options such as pooling computing resources across different innovation centers or using public-cloud or hybrid-cloud systems.
Recalibrate commercial models. New drugs or materials that eventually reach the market typically require high prices per unit to cover the large, fixed costs of experimentation, testing, and regulatory approval over many years. Machine discovery, however, starts to upend this model, as it increases both the pool of successful innovation and compresses the time to market. Industries with “long” innovation cycles will need to move closer to the pricing models of, for example, consumer durables.
Many of the most exciting developments in machine discovery have a strong societal benefit — for example, fighting Alzheimer’s disease or tackling climate change. Businesses will need to work closely with policy makers, scientists, and nongovernmental organizations to ensure that the fruits of machine discovery can be brought to society through appropriate pricing models or cost sharing. More generally, businesses should regard machine discovery not just as a means to increase yields or lower costs, but as a cornerstone of wider social responsibility efforts.
Marcel Proust once said, “The real voyage of discovery consists not in seeking new landscapes, but in having new eyes.” How can machine discovery help your business to better see the innovation possibilities open to it?