KE Xiening

My research interests span four interconnected domains: synthetic chemistry, computational chemistry and physical organic chemistry, rational design of functional materials, and AI for chemistry (AI4C). These are not isolated pursuits but stations along a single intellectual journey, each giving rise to the next.

1. Synthesis: The Joy of Making

I came to chemistry through synthesis, and I stayed for a reason that resists utilitarian justification. A chemical reaction, at its core, is a transformation: one substance becomes another. When we coax atoms along a pathway that nature has never explored, producing a molecule that has never existed in the universe, we experience something close to making in its purest sense. There is a profound sense of agency in this: the chemist as author rather than observer. This is where my interest begins, with free curiosity and the simple, irreducible pleasure of building molecules.

2. Computational Chemistry and Physical Organic Chemistry: Learning to Understand

Synthesis, however, provokes an immediate and uncomfortable question: What, exactly, have we learned from the flask?

We can reproduce a reaction, optimize its yield, and map its scope, but these are empirical facts, not understanding. To truly grasp what happened, we need to descend from the bulk to the molecular: to reconstruct the reaction coordinate, locate the transition state, and disentangle the steric, electronic, and solvent effects that steer the outcome. Computational chemistry, paired with the mechanistic intuition of physical organic chemistry, provides this lens. At the scale of a single molecule, these questions become tractable. We can, with reasonable confidence, explain why a reaction proceeds along one path and not another. The picture is elegant, satisfying, and seductively complete.

3. Rational Design of Functional Materials: The Ambition and the Gap

Emboldened by this molecular-level understanding, one naturally attempts something greater: rational design. If we know how individual molecules behave, surely we can engineer materials, collections of molecules, with prescribed properties. This is the ambition of functional materials chemistry, and at first glance it seems a straightforward extrapolation.

It is not. Success in rational design proves far more elusive than the reductionist logic would predict. We understand the monomer but struggle to predict the polymer; we characterize the ligand but cannot foresee the crystal. Something fundamental is lost in the transition from few to many.

4. AI for Chemistry: Facing More Is Different

The problem, as P.W. Anderson articulated in his landmark essay More Is Different¹, is that quantity itself generates qualitative novelty. Computational chemistry operates comfortably at the scale of tens to hundreds of atoms; experimental chemistry operates at the scale of millimoles, roughly 10²⁰ molecules. The gap between these regimes is not merely a computational inconvenience but a conceptual chasm. Reductionism, Anderson argued, fails here: the behavior of the whole is not a linear superposition of the behavior of its parts. We can never, even in principle, simulate our way from a single-molecule potential energy surface to the emergent properties of a bulk material.

But there is a provocative symmetry worth considering. Large language models perform, in a sense, the reverse operation: they compress trillions of parameters, an astronomically high-dimensional representation of human language, into a few hundred tokens of coherent, context-rich output. Where Anderson identified an irreducibility of scale, modern attention-based architectures² demonstrate a compressibility of scale. This inversion suggests a tantalizing hypothesis: if a transformer can extract the essence from a trillion-dimensional language space, might a similar architecture extract the chemically relevant features from an equally vast molecular and materials space? Could the seemingly irreducible complexity of the “more” be tractable to the right kind of learned compression?

This is the question that animates my interest in AI4C: not merely applying machine learning as a black-box predictive tool, but exploring whether the architecture of attention, the mechanism that collapses complexity into meaning, can bridge the very gap that Anderson identified. It is the attempt to make the more legible again.

References