Four careers, one problem

People occasionally ask why my CV reads like four careers at once: cosmology, then black holes, then physics-informed machine learning, then neutrinos, with quantum computing and financial markets alongside. It is a fair question, and the honest answer is short: I have only ever worked on one problem, and I have carried the same toolkit for it from room to room. What changes each time is the furniture.

The problem is this. You have noisy, incomplete data from a process you understand only partially, and you want a statement about the world that is not merely precise but calibrated: one whose error bar means what it says it means. A small uncertainty is easy to manufacture and worth nothing by itself; what matters is whether the number is honest about how wrong it might be, because somebody is going to act on it. A trader decides how large a position to take. A neutrino collaboration such as DUNE or NOvA decides whether it has seen matter and antimatter behave differently. Neither decision cares how impressive your resolution looked in isolation.

The toolkit, then, is not a set of techniques so much as a set of habits: write the governing equation down before touching the data; decide what you trust and what you merely hope; make the uncertainty a number that survives contact with a decision; and build the software so that all of this runs, rather than merely being true on paper. Those habits are the same in every room, and that is precisely why the rooms can be so different.

It is easier to assert than to see, because each room comes with its own furniture, and the furniture is what makes six projects look like six careers. So here is the tour with the furniture cleared away: one question per room, phrased so that the shared structure is visible.

  • Early-universe cosmology: does a universe that starts out lopsided smooth itself out, and how fast?
  • Black hole magnetospheres: can a spinning black hole act as a particle accelerator, and would we see the gamma rays?
  • Physics-informed neural networks: from a sparse, noisy grid of sensors, can a network that is also taught the governing equation reconstruct the whole wave field in real time, with an error bar you can trust?
  • Neutrinos: when a detector near the beam and one 1,300 km away disagree, how much is physics and how much is the instrument?
  • Quantum-enhanced fluid simulation: when flow through a narrow, branching pipe has to be computed from a 3D scan alone, where does the accuracy actually come from: the solver, or what you assume at the outlets?
  • Markets: when the rules of a market change, how do you notice before you have paid for not noticing?

Metaphorically speaking, six rooms, and the same three tools laid out in each: a governing equation you trust, a measurement you do not entirely, and a decision waiting at the end. In machine-learning terms, the equation is the part of the loss you believe, the data is the part you weight by how far you believe it, and the decision is what calibration is for. What varies between rooms is only which of the three you are least sure of, and therefore which tool you reach for first.

One tool in particular travels better than I expected: the habit of asking where the clean part of the problem ends. In quantum-enhanced fluid dynamics, quantum algorithms for linear systems promise large speed-ups on the interior of the problem, i.e., the solver itself; however, the accuracy of the pipeline is not limited there. It is limited by the outlet boundary conditions: what you assume about the smaller branches downstream that the scan cannot see. Make a thousand-fold faster solver, quantum or classical, and you converge on the wrong answer sooner. In neutrino oscillation fits, the boundary conditions are the beam flux and the interaction cross-sections: several hundred nuisance parameters standing between the equation you trust and the events you count, and with a near detector recording a hundred million interactions, the fit will constrain whatever the model permits it to, whether or not the constraint is physical. In markets, the boundary conditions are impact, inventory, liquidity and latency (how much your own trading moves the price, and how quickly you can get out), and a strategy specified without them has not been specified. Same phrase in three unrelated rooms, and not by coincidence: in each, it names the place where a clean interior problem meets a world you only partly observe, and in each, it is where the real work is.

Once you see this, moving between fields stops feeling like starting over and starts feeling like what it is: carrying the same toolkit into a new room. The furniture is unfamiliar, and for the first few months you bump into it; the tools are familiar, and they are what you are actually there to use. The pattern has held often enough now that I trust it, and it is the reason I no longer find the four-careers question awkward. It is one career, in a house with a lot of rooms.

So that is what this blog is for: the work, from inside whichever room I’m in. What I’m reading; what broke this week; and, now and then, the thing that finally worked. Expect few posts, and real ones, in English, since most of my network is in Europe or the USA and that is the one language that reaches all of it.

Linkedin
Eugene (Yoogeun) Song
Eugene (Yoogeun) Song

Neutrino Physics @ Imperial College London, DUNE & NOvA Collab || Machine Learning || Quantum Computing || Quant Researcher || Multidisciplinary Scientist || Physics × ML × QC × Quant || 🇬🇧 London-based (2023–Present)

Imperial College London | Imperial College London

View profile



Enjoy Reading This Article?

Here are some more articles you might like to read next: