Noether Labs is an AI research lab working to give optical instruments the ability to calibrate, adapt, and recover on their own.
We study how instruments can learn about the physical systems they control: how a telescope should correct light distorted by the atmosphere, which measurement can reveal a drifting calibration, and how an optical receiver should regain a signal after losing it. Our work brings together differentiable physics, sequential measurement, predictive control, and the records needed to understand an instrument’s behavior.
Our long-term goal is to make precision instruments capable of maintaining their own performance as conditions change. We begin with adaptive optics and optical communications, where progress can be measured in residual wavefront error, light delivered to a receiver, and time spent recovering from a disturbance.
Adaptive optics
Atmospheric turbulence changes the shape of a light wave before it reaches an instrument. A wavefront sensor measures the distortion, a controller calculates a correction, and a deformable mirror changes shape to apply it. The atmosphere keeps moving throughout this process. Sensor noise, outdated calibration, and the time taken to compute a command all limit how well the correction works.
In Noether AO, we build differentiable models of wave propagation, pyramid wavefront sensing, and mirror actuation. Differentiation tells us how a change in an optical parameter changes the measured image; this connects the physical model to estimation and control. We compare the model with independent optical references and test classical, predictive, and learned controllers under shared disturbances.
Our ideal pyramid-sensor comparison passed 36 image comparisons, 30 derivative comparisons, and 24 deliberate error controls within a matched finite-grid optical model. The controller question remains open: the learned weights in our retained baseline study did not beat both reference controllers in any of the 12 tested conditions. These outcomes guide what we investigate next.
Learning what to measure
A controller also needs to know whether its model of the instrument is still right. In ProbeBench, we study how an optical system should choose its next calibration measurement when photons and exposures are limited. A deliberate change to the optical phase can reveal information that a single image cannot. The question is which change is worth making, and when another measurement is worth its cost.
Our command-response study compared fixed and adaptive measurement schedules across 24 independent base wavefronts, 288 simulated episodes, and 2,016 exposures. When the applied optical response was 10% lower or higher than the nominal calibration, adaptive probing reduced pooled residual error. Fixed probing performed better under correct calibration and at low initial aberration. The result points toward a specific research problem: deciding when adaptation is useful from measurements available to the instrument itself.
A record of what the instrument did
Learning from an instrument requires more than a stream of images. We need to know when an exposure happened, when its measurement became available, which calibration was in use, and which command the actuator applied. Noether Replay reads adaptive-optics telemetry and reconstructs the evidence available about those events. It reports missing timing, ambiguous command provenance, and calibration gaps so that subsequent analysis can account for them.
Together, these projects form a set of research tools for modeling an instrument, choosing measurements, comparing corrections, and inspecting recorded behavior. We publish source code, experimental protocols, and retained results so that other researchers can examine the assumptions and reproduce the comparisons. Current results cover simulation, numerical optical-model checks, and offline telemetry analysis; validation on physical instruments is a next stage of the work.
From correcting a wavefront to maintaining a link
In optical communications, a corrected wavefront is one part of a larger problem. A receiver must acquire a beam, track it through motion and turbulence, couple light into a fiber, and recover after a fade or loss of signal. Noether Worlds models these stages in a satellite-to-ground optical receiver. Its retained control study compares six approaches across 1,200 reduced-order simulation runs, with a separate diffraction-based study of fiber coupling.
We want to understand how calibration, sensing, and control should work together across those transitions. Our next experiments focus on stronger calibration baselines, tests on previously unseen conditions, and measurements from real instruments. A useful result must survive the sensor noise, actuator limits, and timing constraints of the system it is meant to improve.
Building Noether
Salma Shaik founded Noether in late 2025 and leads its research. The lab is based in Bengaluru. We are building around questions that connect machine learning to physical measurement: when a model can be trusted, how an experiment should be chosen, and what a machine can learn from the consequences of its actions.
We want Noether’s contribution to be original methods, useful research software, and experiments that other people can build on. Our starting point is a focused research program, with room to follow an unexpected result, discard a method that loses to a simpler baseline, and pursue a difficult question before its eventual application is settled.
We welcome research exchange with people working on wave optics, inverse problems, experimental design, control, and scientific instruments. In particular, we are interested in calibration drift, synchronized sensor-and-actuator recordings, and small bench experiments that connect a prediction to a physical measurement. If one of these questions overlaps with your work, write to us.
Read more about the direction of the lab in our founder’s note, explore the research repositories, or follow Noether on X.