Modern artificial intelligence is the product of decades of hard-working scientific research. Now we are beginning to return that effort by accelerating progress across academia.
Since the emergence of AI as a field of research, researchers have dreamed of creating smart tools enough to accelerate humanity's infinite motivation to acquire new knowledge. With the advent of deep learning in the 2010s, this goal has finally become a realistic possibility.
Between 2012 and 2022, the proportion of scientific papers that somehow rely on AI has quadrupled to 9%. Researchers use neural networks to analyze data, conduct literature reviews, and model complex processes across all scientific disciplines. And as technology advances, the scope of issues they can tackle has expanded by that day.
The poster boy for the use of AI in science is undoubtedly the Alphafold of Google Deepmind, whose inventor won the 2024 Nobel Prize in Chemistry. The model used advances in trans (an architecture that enhances large-scale language models) to solve the “protein folding problem” that scientists had been hiding for decades.
The structure of a protein determines its function, but previously the only way to discover its shape was through complex imaging techniques such as X-ray crystallography and cryoelectron microscopy. In comparison, the alphafold was able to predict the shape of the protein, which is nothing but the series of amino acids that make up it.
This has enabled us to predict the shape of all the proteins known to science in just two years. This could have a transformative impact on biomedical research. Released in 2024, Alphafold 3 is even further. It can predict both the structure and interactions of proteins, DNA, RNA, and other biomolecules.
Google also loosened its AI in another area of life science and worked with researchers at Harvard to create the most detailed map of human brain connections to date. The team took ultra-thin slices from a 1-million cubes of the human brain and mapped around 50,000 cells and 150 million synaptic connections using AI-based imaging techniques.
This is the most detailed “connectomb” of the human brain ever produced, with data freely available and providing scientists with key tools for exploring neuronal architecture and connectivity. This can increase understanding of neurological disorders and potentially provide insight into core cognitive processes such as learning and memory.
AI is also revolutionizing the field of materials science. In 2023, Google Deepmind released a graph neural network called GNOME, which predicts 2.2 million new inorganic crystal structures, including 380,000 stable structures that could form the basis of new technologies.
To avoid losing, other big AI developers have also jumped into this space. Last year, Meta released and sourced a dataset with its own transformer-based material discovery model and, importantly, over 110 million material simulations used to train them.
Earlier this year, Microsoft released Mattergen. Mathergen uses the diffusion model, the same architecture used in many image and video generation models, to generate new inorganic crystals. After fine tuning, they showed that they could be encouraged to produce materials with specific chemical, mechanical, electronic, and magnetic properties.
One of the greatest strengths of AI is its ability to model systems that are too complex for traditional computational methods. This is naturally suited for weather prediction and climate modeling, which relies on huge physics simulations currently running on supercomputers.
Google Deepmind's graph cast model was the first to demonstrate the promise of an approach using graph neural networks to generate 10-day predictions with more accuracy than the existing gold standard approach that takes several hours.
AI predictions are so effective that they have already been deployed by the Medium-Range Weather Prediction Center, which was implemented earlier this year with an artificial intelligence forecasting system. The model is faster, 1,000 times more energy efficient, and has increased accuracy by 20%.
Microsoft has created what is called the “Fundamental Model of the Earth System” called Aurora, trained with over a million hours of geophysical data. It is superior to existing approaches in predicting air quality, ocean waves and tropical cyclone paths.
AI also contributes to the fundamental discoveries of physics. When a large hadron collider crushes the particle beam together, millions of collisions occur per second. Sieve through all this data to find interesting phenomena is a monumental task, but now researchers are turning to AI to do it for them.
Similarly, German researchers are using AI via gravitational wave data for indications of neutron star mergers. This will help scientists to detect mergers in time and point to telescopes.
Perhaps most exciting is the promise that AI will take on the role of the scientist itself. By combining lab automation technology, robotics and machine learning, you can now create “autonomous driving labs.” These take high levels of objectives from researchers, such as achieving a specific yield from chemical reactions, and perform the experiment autonomously until it achieves that goal.
Others are going even further, actually involving AI in planning and designing experiments. In 2023, researchers at Carnegie Mellon University showed that Openai's GPT-4 AI “Coscientist” can autonomously plan and implement chemical synthesis of known compounds.
Google has created a multi-agent system equipped with Gemini 2.0 inference model that will help scientists generate hypotheses and propose new research projects. Another “AI Scientist” developed by Sakana AI wrote a machine learning paper that went through the peer review process of the workshop at the prestigious AI conference.
However, because this is all exciting, the AI science acquisition could have potential drawbacks. Neural networks are black boxes that are difficult to decipher internal work, and results are difficult to interpret. And many researchers lack sufficient knowledge of the techniques to capture common pitfalls that can distort results.
Nevertheless, the incredible power of these models remains an important tool to traverse and model data on a scale that goes far beyond human understanding. In smart applications, AI can dramatically accelerate progress in a wide range of fields.
