D$_4$CNN$\times$AnaCal: Physics-Informed Machine Learning for Accurate and Precise Weak Lensing Shear Estimation

Technical Deep Dive: D$_4$CNN$\times$AnaCal

Overview

The D$_4$CNN$\times$AnaCal framework is a physics-informed machine learning approach for accurate and precise weak gravitational lensing shear estimation. Key innovations include:

• A D$4$-equivariant convolutional neural network (D$4$CNN) that encodes the fundamental symmetry properties of galaxy shapes, eliminating second-order biases.
• Integration with the Analytical Calibration (AnaCal) method, which computes the shear response of the model analytically via backpropagation, enabling efficient and accurate bias correction without the need for large simulation suites.
• Demonstration of near-zero multiplicative ($|m| < 10^{-3}$) and additive ($\sim 10^{-5}$) biases across a wide range of observational conditions, meeting the stringent requirements of Stage-IV weak lensing surveys like Euclid, LSST, and Roman.
• $\sim$10% lower shape noise compared to traditional moment-based estimators in the high-noise regime, equivalent to a $\sim$20% gain in effective galaxy number density.

Problem & Context

• Weak gravitational lensing by large-scale structure (cosmic shear) is a premier probe of dark energy and cosmic structure growth, but requires exquisitely accurate shear estimation.
• Key challenges include shape noise, pixel noise, galaxy selection, and point-spread function (PSF) effects, each of which can introduce systematic biases.
• Achieving the sub-percent accuracy required by Stage-IV surveys demands careful control of these biases in galaxy shear estimation.
• Previous machine learning approaches have shown promise, but often fail to encode the fundamental physical symmetries of weak lensing shear.

Methodology

Model Architecture

• The D$_4$CNN architecture hard-codes the rotational and mirror symmetry of galaxy ellipticities into the model design.
• The network applies the same convolutional backbone to eight transformed versions of the input image (the D$_4$ orbit), and combines the resulting features using a weighted average to construct an equivariant representation.
• This equivariant feature is then mapped to the final ellipticity prediction using unbiased odd multilayer perceptrons.

Analytical Calibration

• The model is trained to output ellipticity, but this raw prediction can still contain residual biases.
• The AnaCal method computes the shear response of the model analytically via backpropagation, allowing efficient and accurate bias correction without large simulation suites.
• AnaCal accounts for detection, selection, galaxy model, and PSF anisotropy biases, achieving unbiased shear estimates to second order in shear.

Simulation Setup

• Isolated galaxy images are generated using the GalSim package, mimicking LSST 10-year coadded images.
• Galaxies are distorted by known shear values, and the estimated shear is compared to the true input to quantify bias.
• Orthogonal galaxy pairs are used to cancel image noise and isolate the shape measurement bias.

Results

Accuracy

• The D$_4$CNN$\times$AnaCal framework achieves negligible multiplicative ($|m| < 10^{-3}$) and additive ($\sim 10^{-5}$) biases across a wide range of noise levels, PSF conditions, and magnitude selection cuts, meeting the stringent LSST requirements.
• This level of bias control is maintained even for models trained on as few as 1,000 galaxies, demonstrating the efficiency of the equivariant architecture.

Precision

• The D$_4$CNN model shows $\sim$10% lower shape noise compared to the traditional Fourier Power Function Shapelets (FPFS) estimator in the high-noise regime, equivalent to a $\sim$20% gain in effective galaxy number density.

Interpretation

• The hard-coded D$_4$ symmetry and smooth, differentiable architecture are key to the model's accuracy and efficiency:
  • Equivariance suppresses even-order non-linear terms in the shear response, reducing calibration complexity.
  • Smooth activation functions (like GeLU) preserve gradient continuity, enabling accurate analytical calibration.
• In contrast, models without these properties exhibit larger biases and higher-order non-linearities.
• The framework's modular design allows it to be extended to more powerful backbones (e.g., residual networks, vision transformers) while maintaining the core symmetry and calibration properties.

Limitations & Uncertainties

• The current work focuses on isolated galaxies with known PSF; future extensions must address blending, chromatic PSF effects, and more realistic noise distributions.
• Validation is still required for the impact of PSF model errors and correlated noise on additive biases.

What Comes Next

• Extending the D$_4$CNN$\times$AnaCal framework to blended sources by incorporating scene-native AnaCal calibration.
• Generalizing the equivariant architecture to handle multi-band observations, treating bands as additional input channels.
• Exploring more powerful CNN backbones while preserving the core symmetry and calibration properties.
• Validating the framework's performance under realistic survey conditions, including chromatic PSF effects and correlated noise.