Research atlas Interactive essay · 2026

Finding structure in noisy systems

Three research threads—rare cosmic transients, quantum anomaly detection, and trustworthy quantum infrastructure—connected by a shared problem: choosing the representation that makes a hidden pattern visible.

AuthorJoaquin de Castro
FieldsQuantum · ML · Astrophysics
FormatInteractive research overview

Visualization note. Interactive figures below use synthetic, illustrative data to explain methods and intuition. They are not publication results.

00

Introduction

Noise is not merely an obstacle. It is a clue about which mathematical language a system wants us to use.

Astronomical surveys observe the sky at irregular intervals. Quantum models encode information in complex amplitudes that cannot be read directly. Cryptographic systems must remain secure while their underlying algorithms change. In each case, the raw object is not yet the useful object.

My research asks how to transform these systems without erasing the signal that matters. The figures in this atlas are compact explanations of that process.

01Observeirregular, noisy, incomplete
02Representencode the right invariants
03Decideclassify, detect, adapt
Figure 1. The recurring architecture across the projects in this atlas.

01

Finding rare supernovae in irregular light curves

A telescope does not hand us a smooth movie. It gives us scattered measurements, uneven in time, with uncertainty attached to every point.

Type Ibn supernovae are rare stellar explosions whose light curves can be confused with more common transients. The task is to turn sparse photometry into a representation that a classifier can compare across thousands of objects.

Interactive figure

From scattered observations to a candidate score

illustrative simulation
1Photometryirregular observations
2Gaussian processcontinuous estimate
3Featuresshape, color, timescale
484% candidateclassifier score

What to notice: the fast rise and narrower post-peak evolution remain visible after interpolation, giving the classifier a comparable shape representation.

Why the representation matters

A naive model can mistake missing observations for astrophysical behavior. A Gaussian process instead treats the light curve as a distribution over plausible continuous functions. From that distribution, we can compute features at consistent times and pass them to a tree-based classifier.

10,000+objects processed
+20%+precision improvement
5Ibn candidates surfaced
40 h → <10 minruntime reduction

02

Rewinding quantum models to expose anomalies

An anomaly can be defined as a pattern that a model cannot cleanly undo.

Active galactic nuclei vary over time in ways that are noisy, stochastic, and astrophysically rich. Quantum variational rewinding asks a different question from ordinary classification: after learning a representation of typical behavior, how well can a parameterized quantum process reverse an encoded signal?

Interactive figure

Encode → evolve → rewind → score

current research direction

Classical signal

Variational circuit

0.18 rewind error

Interpretation

Consistent with learned dynamics

The circuit approximately returns the encoded state toward its reference, producing a low anomaly score.

The research question

The important question is not whether a quantum circuit can draw a complicated boundary. It is whether the circuit learns a representation of temporal structure that is useful under realistic noise, data volume, and classical baselines. The project therefore treats quantum advantage as an empirical hypothesis—not an assumption.

03

Designing quantum systems people can trust

A powerful algorithm is not yet a dependable system. Trust comes from knowing what can change, what must remain invariant, and how failure is contained.

My work in quantum information and post-quantum security spans blind quantum computing, error-correction ideas, random-circuit simulation, and crypto agility. These topics share a concern with structure under changing conditions.

Interactive figure

Resilience is an architecture, not an algorithm

01

Separate interface from implementation

Systems adapt faster when the surrounding workflow does not depend on one algorithm forever.

02

Measure what survives intervention

A representation is more convincing when it remains meaningful under ablation, noise, or substitution.

03

Make uncertainty visible

Trustworthy technical work exposes assumptions, baselines, and failure modes rather than hiding them.

04

The throughline

Find the representation that preserves what matters—and makes the next decision possible.

This is the thread connecting my research and the systems I want to build: scientific models that reveal hidden structure, infrastructure that can adapt without breaking, and institutions that help more people work at the frontier.