The first time I laid eyes on an iPhone, a thought thundered through my mind: “This can’t be real; we’ve stepped into the future.” Yet it was real indeed, and it took me—us, really—only a few days to acclimate to its existence. Quite astonishing, isn’t it? This sense of disbelief mirrors my feelings towards machine learning and AI. We’ve become accustomed to a world where large language models and other AI systems enhance our everyday lives, a scenario that equally applies to science. We’re at a crucial juncture, a tipping point, where change occurs at breakneck speed. In what directions will this lead us? Just a few days ago, during one of our conversations on LinkedIn, we were dissecting the role of simulations and AI in scientific discovery. In that discussion, a bold reader suggested that by 2026, AI might replace both traditional experiments and simulations. Agree or disagree, this idea highlights a significant challenge: scientific understanding. I can concede that, eventually, we may possess AI capable of providing the correct answers, but how do we train such AI without data from simulations or experiments? And more critically, how will we understand the mechanisms and principles behind these answers? This brings us straight to today’s story: how is the concept of scientific understanding evolving in the age of AI? What exactly does this entail, and why does it matter? Following Krenn’s recent Nature review, we are going to seek the answers together. Shall we begin?
Beyond Discovery: Scientific Understanding
Picture a digital oracle, shrouded in mystery, capable of forecasting every experiment or chemical reaction with uncanny precision. Meet Laplace’s AI-demon, a character we’ve toyed with before in our philosophical musings on From Atoms To Words.
The thought of such an oracle might send shivers down your spine, I know. Still, when you consider the advancements in machine learning and AI technologies, like AlphaFold’s breakthroughs in protein folding or the sophisticated large language models for chemistry—the concept doesn’t seem so far-fetched anymore. These tools might be far from being a digital oracle, but they are already handing us answers on a silver platter, signaling a disruption in how we approach science.
And yet, despite the potential abundance of answers, we find ourselves scratching our heads over the underlying mechanisms. How do these technologies yield their outcomes? What principles allow them to work so effectively? What do their answers truly mean?
A phenomenon P can be understood if there exists an intelligible theory T of P such that scientists can recognize qualitatively characteristic consequences of T without performing exact calculations.
de Regt and Dieks | 2005
After all, “true science” is more than just achieving meaningful results; it’s a journey towards understanding fundamental processes, through theories and the ability to qualitatively predict without performing exact calculations. While having the correct answers marks scientific discovery, bypassing the journey towards scientific understanding leaves our knowledge incomplete.
Discovery alone is not enough for science to progress. It needs scientific understanding, the cornerstone of human inquiry. It’s what drives us to look beyond, to grasp the how and the why of natural phenomena.
The distinction between scientific discovery and understanding is subtle yet profound. So, inspired by Krenn’s recent Nature review, let’s explore six concrete examples that vividly demonstrate this dichotomy.
Curious to dive in?
More on From Atoms To Words:
▸ ReaxFF Molecular Dynamics: Simulating Complexity Beyond Quantum Chemistry
▸ How Can Coarse-Grained Simulations Reveal Geckos’ Wall-Clinging Skills?
▸ The Evolution of Quantum Chemistry: From Pencil and Paper to Quantum Computing
Discovery without Scientific Understanding: Six Concrete Examples
We’re about to dive into some thought-provoking cases that illustrate how we might find answers but lack scientific understanding. Let’s pull back the curtain and take a closer look, shall we?
- Screening Organic Laser Diodes: Imagine combing through a crowd of 1.6 million to find the one—the molecule that shines brighter and more efficiently than all its peers for organic laser diodes. Machine learning achieved just that, identifying molecules with stellar quantum efficiency. But here’s the kicker: even with this technological marvel, we’re left pondering why these molecules perform so well.
- Predicting Protein Structures: Then there’s AlphaFold—a groundbreaking tool that has revolutionized the prediction of protein structures. But, and it’s a big but, it operates like a black box: predictions come out spot on (most of the time), yet how it unravels the complexities of protein structure remains elusive.
- Uncovering Crystal Structures: Talking about data-rich, think of GNoME, an AI model that has uncovered over two million new crystal structures. Impressive, right? But when it comes down to it, making sense of this colossal data set still feels like finding a needle in a haystack. Yes, we have a mountain of data at our fingertips, but we’re still fumbling in the dark about what it all means. It’s a strong case of more information not necessarily translating into more scientific understanding.
- Rediscovering the Heliocentric Model: Iten and team crafted a neural network inspired by human reasoning, able to unearth foundational concepts such as Copernicus’ heliocentric theory directly from experimental data—all without relying on any preconceived notions about the system under investigation. While this is indeed pretty awesome, can AI deepen our scientific understanding?
- Revealing the Arrow of Time: By training an AI to discover the arrow of time, Seif and team not only nodded to the fluctuation theorem but also highlighted AI’s talent for pinpointing entropy as a key player. And again, the question remains: will an AI that can rediscover physical laws also contribute to new scientific understanding? Can it produce original hypotheses?
- Parsing Cosmic Equations of Motion: And then there’s the tale of using machine learning to parse the laws governing our solar system. Lemos and team employed a graph neural network trained on 30 years of solar system trajectory data to automatically deduce Newton’s law of gravitation. This success story confirms AI’s ability to retrieve known laws from data, but can AI generate new knowledge?
I agree with Krenn when they argue that these examples lay bare a crucial dilemma: the real challenge isn’t about discovery or even rediscovery—it’s about charting a course to novel insights. Simply put: can AI advance our scientific understanding?
Echoing our musings on computational creativity and serendipity, gaining new understanding from advanced computational systems means discovering new ideas, principles, concepts or even theories that scientists can leverage in various scenarios.
The journey from discovery to scientific understanding is complex, no doubt. It calls for a synergy between AI’s computational might and our insatiable human curiosity for knowledge.
How can this be achieved? How do we transition from data-driven discovery to computer-assisted scientific understanding?
More on From Atoms To Words:
▸ Is Machine Learning Going to Replace Computational Chemists?
▸ Bridging Theory and Experiment: 14 Reasons Chemical Simulations Stand as the Third Pillar of R&D
▸ Large Language Models for Chemistry: Is the Beginning of a New Era?
The Three Dimensions of Computer-assisted Scientific Understanding
Imagine a world where advanced computational systems light our way through the mysteries of science. Sounds like a sci-fi dream, doesn’t it? Well, Krenn has sketched out a roadmap in the form of dimensions that demonstrates—at least in principle—how simulations and AI may advance our scientific understanding. Let’s explore what such a visionary framework entails.
▸ Dimension 1: The Computational Microscope
Remember the way microscopes revolutionized biology by bringing the tiny world of bacteria into view? In a similar vein, computational microscopes can investigate complex biological, chemical, and physical processes that escape the naked eye, working on scales or under conditions that conventional experimentation can’t easily probe.
For example, we’ve discussed some jaw-dropping cases of the computational microscope, such as the beautiful work by Casalino on the SARS-CoV-2 virus. Their all-atom molecular dynamics simulations revealed hidden biological functions of the spike protein, transforming our scientific understanding of glycans in biological systems. This is the power of computational microscopes—they not only shift our perspective but also ignite the spark for new ways to probe the machinery of nature, bit by bit.
The cases of computational microscopes we’ve talked about in previous stories confirm that our computational present looks bright. But what about the future? Well, it may be even brighter.
With advancements in hardware (think quantum computing), alongside smarter algorithmic strategies and richer data representations, the capabilities of computational microscopes are only set to soar. Imagine immersing yourself in augmented 3D environments or watching time-lapse videos that give you access to hard-to-detect phenomena. These tools are poised to make scientific understanding more intuitive, more profound, and, yes, even more thrilling.
▸ Dimension 2: AI as a Resource of Inspiration
A whopping 70 years ago, Alan Turing himself was caught off guard by how often machines threw him for a loop, remarking, “Machines take me by surprise with great frequency.” If you think about it, serendipity and creativity are the foundation of advances in science.
So, we’re led to wonder: could computer algorithms inspire such ideas systematically, thereby accelerating scientific understanding and technological progress?
According to Krenn, the answer is a resounding yes; they characterize this aspect of AI as a Resource of Inspiration, suggesting that AI could be more than just a tool enhancing our ability to compute and discover. It may serve as a muse that inspires scientists to push beyond the boundaries of known paradigms.
And honestly, seeing how today’s AI unravels complex patterns and extracts meaningful insights really proves Krenn’s point. For those who’ve skimmed through the pages of From Atoms To Words, you’re well aware of the seismic shifts machine learning and AI have triggered across the scientific spectrum, from drug discovery to chemistry and materials science.
An oracle that correctly predicts the outcome of every particle physics experiment, the products of every possible chemical reaction or the function of every protein would revolutionize science and technology. However, scientists would not be entirely satisfied because they would want to comprehend how the oracle made these predictions.
Mario Krenn | 2022
But the analytical strength of AI doesn’t stop at the edge of the digital or the confines of the laboratory. It dives deep into the ocean of scientific literature, mining gems of knowledge and bridging disciplines with its discoveries. The sheer volume of publications AI can sift through has not only led to a renaissance of core scientific concepts but has also unveiled previously unknown structure-property relationships.
And let’s not overlook AI’s ability to translate its findings into formats we can better comprehend, such as mathematical formulas. This skill has enabled scientists to develop new theories and models across an array of fields, from the mechanics of the cosmos to the intricacies of quantum physics and beyond, emphasizing AI’s potential role in advancing scientific understanding.
These and further approaches like artificial curiosity and computational creativity—topics we’ve recently explored—promise to be the catalysts of scientific innovation, guiding human scientists in their relentless goal of deciphering the fundamental workings of nature.
▸ Dimension 3: AI as the Agent of Understanding
Up to this point, we’ve looked at some pretty mind-boggling ideas. But get ready, because we’re about to take our speculations to the next level. So, picture algorithms not just crunching numbers in solitude but stepping up as teachers, imparting new scientific insights directly to us, mere humans. This would mark the threshold where AI surpasses its traditional role, transitioning into what Krenn has dubbed the Agent of Understanding.
Yes, you read that right. Krenn envisions a within-our-lifetime future where machines evolve from predictors to enlighteners, ushering us into the age of ultrastrong machine learning. These are machines that, in the definition of AI pioneer Donald Michie, can illuminate the path of knowledge for humans.
But how do we know when AI has truly stepped into its role as an agent of understanding? Krenn lays out two critical milestones:
- The Mastery of Theoretical Insights: AI-driven insight goes beyond data analysis to embrace a profound comprehension of the underlying theories.
- The Power to Enlighten: AI’s real strength manifests in its ability to transcend the black-box phase and communicate its insights, making complex concepts understandable to us.
How, then, do we assess these ambitious milestones?
With the scientific understanding test, a concept that might remind you of the Turing and Feigenbaum tests. This benchmark challenges AI not just to match but to mirror the teaching efficacy of a human expert, marking a pivotal step towards an intelligent machine that contributes as meaningfully to our collective knowledge as any human scientist.
More on From Atoms To Words:
▸ Can Quantum Chemistry Simulations Help Trace the Origin of Life?
▸ Curiosity, Ingenuity, Persistence – Andre Geim’s Random Walk to the Discovery of Graphene
▸ When Will RNA Structure Prediction Get Its AlphaFold Breakthrough?
A final personal touch
Lately I have been venturing deeper into the whys behind science, stepping into a philosophical arena that stretches my comfort zone. And perhaps yours.
There was a time when I thought philosophy would wane in relevance as scientific understanding expanded. You might recall Stephen Hawking’s bold declaration in his book The Grand Design: “Philosophy is dead.”
Let’s just say that our Stewie was, ahem, stirring the pot.
I believe that in our current era, where data-driven insights, whether through simulations, machine learning, or AI, are becoming central, the disciplined curiosity of philosophy proves more essential than ever. Are simulations theory or experiments? Are AI environments a simulation? What’s the role of serendipity in the age of AI? That’s precisely why I engage with works like those from Krenn—to reflect, to pause, to learn. And is there a better way to learn than to write about it?
So, what have we learned today?
Science is on the move. Discovery is increasingly reachable. Concepts such as intelligence, serendipity, and scientific understanding are blossoming into a colorful garden of meanings.
In facing this evolution, my strategy is, yes, to reflect, to pause, to learn. But also to remain open-minded, especially when presented with statements as bold as the one from Krenn: “We firmly believe that these research efforts can—within our lifetimes—transform AI into true agents of understanding that will directly contribute to one of the main goals of science, namely, scientific understanding.”
The ultimate takeaway? This evolution isn’t on its way. It’s happening before our very eyes. And our responsibility? To gear up for what’s coming.
If you enjoyed this dive into discovery and scientific understanding, I’d love to hear your thoughts. Agree, disagree, or have a totally wild theory of your own? Let’s connect! Subscribe to my LinkedIn newsletter and let’s keep the conversation rolling.