Google's AI Breakthroughs: Advancing Scientific Frontiers in 2024

Here's How AI Helped Google Make Notable Scientific Discoveries in 2024

This was the inaugural edition of the AI for Science Forum event Google has revealed notable scientific breakthroughs made this year that were possible due to advancements in artificial intelligence (AI) technology. On Monday, Google DeepMind co-hosted the inaugural edition of the AI for Science Forum in London alongside the Royal Society. During the event, the Mountain View-based tech giant recapped achievements such as using an AI model to predict protein structures, expanding its flood forecasting system, and wildfire detection and tracking system. DeepMind was also able to build a system that is capable of controlling plasma with a nuclear fusion reactor. The tech giant stated that AI has played a key role in the ongoing year in solving many confounding problems in science using its computational techniques. The company also emphasised that AI is not a replacement for scientists but can become a crucial assistant for them. One of the biggest achievements of Google DeepMind was when the AI research wing of the tech giant used its AlphaFold 2 AI model to predict structures of 200 million proteins. The company highlighted that this discovery pushed the scientific community decades ahead as the determination of the 3D structure of a single protein can take up to a year. Notably, Demis Hassabis and John Jumper, the individuals behind the project were awarded the Nobel Prize in Chemistry 2024 for this discovery. Google also partnered with Lichtman Lab at Harvard to map a piece of the human brain of an unprecedented level of detail. This project was released this year and revealed structures within the human brain that were previously unseen. In 2024, AI also helped Google improve its prediction and tracking systems. The company's riverine flood forecasting system was expanded at scale in 2024 and now covers 100 countries and 700 million people globally. The tech giant also partnered with the US Forest Service to develop the FireSat AI model that can detect and track wildfires as small as the size of a classroom within 20 minutes. GraphCast, a machine learning research model developed by Google DeepMind, can now predict the tracks of cyclones. The company claims it can predict such weather-based disturbances faster and more accurately compared to traditional weather simulation systems. Advances were also made in mathematical reasoning and quantum computing. DeepMind's AlphaGeometry AI system, which was launched in 2024, solved complex geometry problems at a level similar to a human Olympiad gold medallist. Google researchers also worked with UC Berkeley and Columbia University to perform chemical simulations on a quantum computer to predict chemical reactivity and kinetics. Looking towards sustainable energy, the tech giant announced the Graph Networks for Materials Exploration (GNoME) which discovered 3,80,000 materials that are stable at low temperatures, opening new avenues to make better solar cells, batteries, and potential superconductors. The tech giant also made breakthroughs in nuclear fusion, which is considered the energy of the future. Collaborating with the Swiss Plasma Center at the Swiss Federal Institute of Technology Lausanne, Google DeepMind announced the development of an AI system that can control the plasma inside a nuclear fusion reactor without any manual assistance. This is still a work in progress, but the company said it is a critical step towards stable fusion and abundant clean energy.

9 ways AI is advancing science

1. Cracking the 50-year "grand challenge" of protein structure prediction Experts have described demystifying protein folding as a "grand challenge" for decades. In 2022, Google DeepMind shared the predicted structures of 200 million proteins from their AlphaFold 2 model. Previously, determining the 3D structure of a single protein typically took a year or more -- AlphaFold can predict these shapes with remarkable accuracy in minutes. By releasing the protein structure predictions in a free database, this has enabled scientists around the world to accelerate progress in areas like developing new medicines, fighting antibiotic resistance and tackling plastic pollution. As a next step, the AlphaFold 3 model builds on AlphaFold 2 to predict the structure and interaction of all of life's molecules. Few things have held more mystery throughout time than the human brain. Developed over 10 years of connectomics research, Google partnered with others, including the the Lichtman Lab at Harvard, to map a tiny piece of the human brain to a level of detail never previously achieved. This project, released in 2024, revealed never-before-seen structures within the human brain. And the full dataset, including AI-generated annotations for each cell, has been made publicly available to help accelerate research. When Google's flood forecasting project began in 2018, many believed it was impossible to accurately deliver flood forecasting at scale, given the scarcity of data. But researchers were able to develop an AI model that achieves reliability in predicting extreme riverine events in ungauged watersheds at up to a five-day lead time with reliability matching or exceeding that of nowcasts (zero-day lead time). In 2024, Google Research expanded this coverage to 100 countries and 700 million people worldwide -- and improved the AI model so it offers the same accuracy at a seven-day lead time as the previous model had at five. Wildfires are increasingly upending communities around the world due to hotter and drier climates. In 2024, Google Research partnered with the U.S. Forest Service to develop FireSat, an AI model and new global satellite constellation designed specifically to detect and track wildfires the size of a classroom by providing higher-resolution imagery within 20 minutes. This will allow fire authorities to respond more quickly, potentially saving lives, property and natural resources. In 2023, Google DeepMind launched and open sourced the model code for GraphCast, a machine learning research model that predicts weather conditions up to 10 days in advance more accurately and much faster than the industry gold-standard weather simulation system (HRES). GraphCast can also predict the tracks of cyclones (and associated risks like flooding) with greater accuracy, and accurately predicted Hurricane Lee would hit Nova Scotia three days before traditional models. AI has always struggled with complex math due to a lack of data and reasoning skills. Then, in 2024, Google DeepMind announced AlphaGeometry, an AI system that solved complex geometry problems at a level approaching a human Olympiad gold-medalist -- a breakthrough in AI performance and the pursuit of more advanced general AI systems. The subsequent Gemini-trained model, AlphaGeometry 2, was then combined with a new model AlphaProof, and together they solved 83% of all historical International Mathematical Olympiad (IMO) geometry problems from the past 25 years. In demonstrating AI's growing ability to reason, and potentially solve problems beyond current human abilities, this moved us closer to systems that can discover and verify new knowledge. Google researchers worked with UC Berkeley and Columbia University to perform the largest chemistry simulations to date on a quantum computer. The results, published in 2022, were not only competitive with classical methods, but they also did not require the burdensome error mitigation typically associated with quantum computing. The ability to conduct these simulations will offer even more accurate predictions of chemical reactivity and kinetics, which is a precursor for applying chemistry in new ways to help solve real-world challenges. In 2023, Google DeepMind announced Graph Networks for Materials Exploration (GNoME), a new AI tool that has already discovered 380,000 materials that are stable at low temperatures, according to simulations. At a time when our world is looking for new approaches to energy, processing power and materials science, this work could pave the way to better solar cells, batteries and potential superconductors. Plus, to help this technology benefit everyone, Google DeepMind made GNoME's most stable predictions available via the Materials Project on their open database. As the old joke goes, "Fusion is the energy of the future -- and it always will be." Controlling and using the energy that fuels stars -- including our own sun -- has been beyond the realm of science. Then in 2022, Google DeepMind announced that it developed AI that can control the plasma inside a nuclear fusion reactor autonomously. By collaborating with the Swiss Plasma Center at EPFL, Google DeepMind built the first system capable of autonomously stabilizing and shaping the plasma within an operational fusion reactor, taking a critical step toward stable fusion and abundant clean energy for everyone.

Demis Hassabis-James Manyika: AI will help us understand the very fabric of reality

But in the course of that journey you will also discover that, despite all this incredible progress, there are surprising limits to the things we know. We are still nowhere near answering some of the biggest questions, like the nature of time, consciousness, or the very fabric of reality. To make progress towards answering these profound questions, new tools and approaches will almost certainly be needed. Artificial intelligence (AI) is one such tool, and we've always believed that it could, in fact, be the ultimate tool to help accelerate scientific discovery. We've been working toward this goal for more than 20 years. DeepMind (now Google DeepMind) was founded with the mission of responsibly building Artificial General Intelligence (AGI), a system that can perform almost any cognitive task at a human level. The immense promise of such systems is that they could then be used to advance our understanding of the world around us, and help us solve some of society's greatest challenges. In 2016, after we'd developed AlphaGo, the first AI system to beat a world champion at the complex game of Go, and witnessed its famously creative Move 37 in Game 2, we felt the techniques and methods were in place to start using AI to tackle important open problems in science. At the top of that list was the 50-year-old grand challenge of protein folding. Proteins are the building blocks of life. They underpin every biological process in every living thing, from the fibers in your muscles to the neurons firing in your brain. Each protein is specified by its amino acid sequence (roughly its genetic sequence) and spontaneously folds into a three-dimensional structure. The shape of a protein is important because it tells you a lot about what the protein does -- information that's critical for things like understanding diseases, and drug discovery. Predicting the 3D shape of a protein directly from its 1D amino acid sequence is known as the "protein folding problem." It's incredibly challenging because there are estimated to be more potential ways that an average protein can fold than there are atoms in the universe. Finding a protein's structure experimentally can take years of painstaking and expensive work. It can typically take a grad student their entire PhD to produce just one structure. After a monumental collective effort spanning decades, structural biologists had determined around 170,000 of these structures and deposited them in the Protein Data Bank (PDB). AlphaFold was our solution to this problem -- and it was recognized with this year's Nobel Prize in Chemistry. AlphaFold learns a complex model of proteins from the structures in the PDB and other related data. It can then, in minutes, predict the structure of a novel protein down to atomic accuracy (i.e., to within the width of an atom on average). As AlphaFold is so fast as well as accurate, over the course of a year, we were able to use it to predict the structure of nearly every protein known to science: over 200 million proteins -- a task that would have taken approximately a billion years of PhD time. To have the most beneficial impact on society, we made AlphaFold and all of its predicted structures freely and openly available for anyone in the world to use, in partnership with the European Bioinformatics Institute (EMBL-EBI). In just three years, over 2 million researchers from 190 countries have used it to advance their important work, from designing enzymes to tackle plastic pollution to creating a molecular syringe capable of delivering therapeutic proteins directly into human cells to developing effective malaria vaccines to combating antimicrobial resistance, and much more. We established Isomorphic Labs to further build on these breakthroughs and use AI to revolutionize the drug discovery process, making it faster and less expensive. This is what we call: science at digital speed. Indeed, with AI as a tool, scientists are making great progress in nearly every field of scientific endeavor. At Google DeepMind and Google Research, working with academic collaborators, we have been using AI to help control the shape of plasma in a fusion reactor, discover faster matrix multiplication algorithms, make mathematical discoveries, discover new materials, explore quantum dynamics, understand behaviors in the brain, draft the first reference pangenome, advance the synaptic-level mapping of the human brain, and make better weather predictions. Advances like these are starting to have a very real, beneficial impact on people's lives. For example, flood prediction is becoming a more frequent and urgent problem due to climate change. Yet only a small percentage of the world's rivers have streamflow gauges that can provide direct forms of early warning. Using publicly available data, we used AI to accurately predict riverine flooding up to seven days in advance. Scaling up from an initial pilot in Bangladesh, our early-warning Flood Hub platform now covers hundreds of millions of people in over 80 countries around the world, including in vulnerable and data-scarce regions. Of course, as we pursue bold leaps in scientific progress, we must also embrace our collective responsibility to build AI in a way that benefits humanity and mitigates against potential harms and misuse. Alongside scientists and technologists, we have to ensure philosophers, ethicists, social scientists, and national scientific academies are brought into the conversation about the future of AI. A safe and prosperous future with AI is possible only if industry works closely together with government, academia, and civil society to chart the way forward. This includes work towards a regulatory framework that fosters innovation and advances AI-enabled opportunities that benefit everyone. AI will be one of the most transformative technologies ever invented. We must approach it with the seriousness and respect it deserves. Although there are many challenges to overcome, both technical and ethical, we believe that with enough time and care, human ingenuity will solve them. We have to be both bold and responsible. As AI accelerates the pace of progress itself, new discoveries will build on each other in a virtuous cycle. We may very well be on the threshold of a new golden age of discovery, one that brings us closer than ever to understanding some of the deepest mysteries of the universe, and our place in it.