betaFeedback

WorksPhysics and Astronomy

Unexplored Opportunities for Automatic Differentiation in Astrophysics

Abstract

We present a systematic analysis of automatic differentiation (AD) applications in astrophysics, identifying domains where gradient-based optimization could provide significant computational advantages. Building on our previous work with GRAF (Gradient-based Radar Ambiguity Functions), which discovered optimal radar waveforms achieving 4x computational speedup by exploring the trade-off space between conflicting objectives, we extend this discovery-oriented approach to astrophysical parameter spaces. While AD has been successfully implemented in several areas including gravitational wave parameter estimation and exoplanet atmospheric retrieval, we identify nine astrophysical domains where, to our knowledge, gradient-based exploration methods remain unexplored despite favorable mathematical structure. These opportunities range from discovering novel solutions to the Einstein field equations in exotic spacetime configurations to systematically exploring parameter spaces in stellar astrophysics and planetary dynamics. We present the mathematical foundations for implementing AD in each domain and propose GRASP (Gradient-based Reconstruction of Astrophysical Systems & Phenomena), a unified framework for differentiable astrophysical computations that transforms traditional optimization problems into systematic exploration of solution spaces. To our knowledge, this is the first work to systematically delineate unexplored domains in astrophysics suitable for automatic differentiation and to provide a unified, mathematically grounded framework (GRASP) to guide their implementation.

A Preprint

© 2025 Marc Bara – Released under CC BY 4.0

Patent Pending – U.S. Provisional Patent Application No. 63/842,854

Introduction

The application of automatic differentiation (AD) to scientific computing has emerged as a transformative methodology for parameter estimation and optimization problems. Originally developed for neural network training, AD frameworks such as PyTorch [1], JAX [2], and TensorFlow [3] provide exact gradient computation through computational graphs, enabling efficient optimization in high-dimensional parameter spaces.

In the context of signal processing applications, we previously demonstrated how AD transforms optimization problems into systematic exploration tools. Our GRAF (Gradient-based Radar Ambiguity Functions) system achieved a 4×4 \times reduction in computation time by discovering optimal radar waveforms through exploration of the trade-off space between peak sidelobe levels and low probability of intercept characteristics [4]. Crucially, GRAF demonstrated that gradient-based methods not only optimize faster but discover superior solutions in multi-objective spaces—waveforms that no engineer would have designed yet outperform traditional approaches on all metrics simultaneously. This success in discovering optimal configurations within complex parameter spaces motivated our investigation into similar opportunities in astrophysics.

From Optimization to Discovery

The key insight from GRAF extends beyond computational efficiency: gradient-based exploration of parameter spaces discovers solutions that elude human intuition. In radar waveform design, the joint optimization of conflicting objectives (detection performance vs. stealth characteristics) revealed Pareto-optimal solutions in regions of parameter space that traditional design approaches never explore. This paradigm shift—from optimizing human-conceived designs to discovering the full landscape of optimal solutions—motivates the extension to astrophysical domains.

In astrophysics, many fundamental problems involve exploring vast parameter spaces subject to physical constraints:

  • What spacetime geometries minimize exotic matter while permitting superluminal motion?

  • Which stellar models simultaneously satisfy asteroseismic frequencies and spectroscopic constraints?

  • What antenna gain configurations optimize interferometric calibration across multiple observing conditions?

These questions require not just optimization of known solutions, but systematic discovery of all possible solutions within the constraints of physics.

The Power of Automatic Differentiation: From Forward Models to Discovery

The transformative insight underlying both GRAF and GRASP is deceptively simple: if we can compute something, automatic differentiation lets us optimize it through iterative refinement. This principle, familiar from neural network training, extends powerfully to physical systems.

Consider the current workflow in astrophysics when dealing with complex, non-convex optimization landscapes:

  1. Forward model: Given parameters θ\theta, compute observables (e.g., given stellar mass, compute oscillation frequencies)

  2. Comparison: Evaluate how well the model matches data or satisfies constraints

  3. Population-based search: Use genetic algorithms, particle swarm, or similar methods that require hundreds to thousands of forward evaluations to estimate good parameter updates

These methods work but are computationally expensive because they treat the forward model as a black box, learning about the landscape only through repeated sampling.

Automatic differentiation transforms this into an efficient discovery process:

  1. Forward model: Same as before - compute observables from parameters

  2. Loss function: Quantify the mismatch or constraint violation as L(θ)\mathcal{L}(\theta)

  3. Backward pass: AD automatically computes ∇θL\nabla_\theta\mathcal{L} - the exact direction to improve

  4. Parameter update: θnew=θold−η∇θL\theta_{new} = \theta_{old} - \eta\nabla_\theta\mathcal{L}

  5. Iterate: Repeat until convergence

The crucial advantage: while genetic algorithms require hundreds of forward evaluations to estimate a single good update direction, AD provides the exact gradient at a cost comparable to just one or two forward passes. This represents a 100100–1000×1000\times reduction in computational cost per iteration.

This efficiency enables a paradigm shift. Rather than being limited by computational budget to explore only small regions of parameter space, we can:

  • Start from many random initializations to explore diverse regions efficiently

  • Navigate through million-dimensional parameter spaces (e.g., neural network weights)

  • Discover solutions in regions traditional methods never reach due to computational constraints

  • Find optimal trade-offs between competing objectives in minutes instead of days

Just as gradient descent in neural networks discovers feature representations no human would design, gradient-based exploration in physics discovers configurations no scientist would conceive—yet they satisfy all physical laws and optimize our objectives.

Technical Foundations of Automatic Differentiation

While the discovery paradigm drives our approach, we briefly review the mathematical machinery that makes it possible. AD computes exact derivatives by decomposing functions into elementary operations and applying the chain rule systematically.

Forward Mode: For a function f:Rn→Rmf : \mathbb{R}^n \to\mathbb{R}^m, forward mode computes directional derivatives:

y˙=∂f∂xx˙(1)\dot{y} = \frac{\partial f}{\partial x}\dot{x} \tag*{(1)}

where x˙\dot{x} is a seed vector. This is efficient when n≪mn \ll m.

Reverse Mode: More relevant for optimization where typically m=1m = 1 (scalar loss), reverse mode computes:

xˉ=(∂f∂x)Tyˉ(2)\bar{x} = \left(\frac{\partial f}{\partial x}\right)^{\mathrm{T}}\bar{y} \tag*{(2)}

where yˉ\bar{y} is the adjoint. This requires one forward pass and one backward pass, regardless of nn.

Complex Differentiation: For complex-valued functions, the Wirtinger derivatives provide:

∂f∂z=12(∂f∂x−i∂f∂y),∂f∂z∗=12(∂f∂x+i∂f∂y)(3)\frac{\partial f}{\partial z} = \frac{1}{2}\left(\frac{\partial f}{\partial x} - i\frac{\partial f}{\partial y}\right), \qquad\frac{\partial f}{\partial z^*} = \frac{1}{2}\left(\frac{\partial f}{\partial x} + i\frac{\partial f}{\partial y}\right) \tag*{(3)}

where z=x+iyz = x+iy. For holomorphic functions, ∂f∂z∗=0\frac{\partial f}{\partial z^*} = 0, recovering standard complex differentiation. This formalism is essential for radio interferometry and other complex-valued problems where functions are generally not holomorphic.

Current State of AD in Astrophysics

To our knowledge, automatic differentiation has already proved useful in a diverse set of astrophysical workflows, including gravitational–wave parameter estimation [5, 6], strong–lensing mass modelling [7], cosmological parameter inference [8], exoplanet spectroscopic retrieval [9], pulsar–timing residual analysis [22], transit–photometry fitting [10], fast–radio–burst dispersion modelling [23], Zeeman–Doppler imaging of stellar magnetic fields [24], and exploratory optimisation of wormhole geometries [25]. Collectively, these implementations confirm that gradient–based methods are both feasible and advantageous whenever the underlying physics is differentiable and the search space is high–dimensional—precisely the regime targeted by GRASP.

Despite these successes, a substantial swath of astrophysical problems with comparable differentiable structure remains unexplored. In this paper we delineate nine such frontiers and provide the mathematical scaffolding required to equip each of them with gradient-based optimisation.

Mathematical Framework for Astrophysical Applications

General Discovery Framework

Most astrophysical problems can be formulated as exploration of solution spaces subject to constraints:

S∗={θ:C(θ)=0,L(θ)≤Lmax⁡}(4)\mathcal{S}^* = \{\theta: \mathcal{C}(\theta) = 0, \mathcal{L}(\theta) \le\mathcal{L}_{\max}\} \tag*{(4)}

where S∗\mathcal{S}^* represents the set of all valid solutions, C(θ)\mathcal{C}(\theta) represents physical constraints (e.g., Einstein equations, conservation laws), and L(θ)\mathcal{L}(\theta) represents objectives to minimize (e.g., exotic matter, fitting residuals).

The discovery process involves:

  1. Exploration: Systematically sampling the parameter space θ\theta

  2. Constraint satisfaction: Ensuring physical validity through differentiable constraints

  3. Pareto frontier identification: Finding trade-offs between competing objectives

  4. Solution characterization: Identifying qualitatively different solution families

This transforms single-point optimization into comprehensive mapping of solution landscapes.

Key Differentiable Operations in Astrophysics

Several mathematical operations appear repeatedly across astrophysical applications:

Integration: Many forward models involve integrals

I(θ)=∫Ωf(x;θ) dx(5)I(\theta) = \int_{\Omega} f(x;\theta)\,dx \tag*{(5)}

These can be differentiated using Leibniz’s rule when the integrand is smooth.

Differential Equations: Stellar structure, orbital dynamics, and field evolution often involve ODEs/PDEs:

dydt=f(y,t;θ),y(0)=y0(θ)(6)\frac{dy}{dt} = f(y,t;\theta), \qquad y(0) = y_0(\theta) \tag*{(6)}

The sensitivity equations or adjoint methods provide gradients.

Fourier Transforms: Essential for interferometry and time series analysis:

f~(k)=∫f(x)e−ikx dx(7)\tilde{f}(k) = \int f(x)e^{-ikx}\,dx \tag*{(7)}

The FFT is differentiable, enabling end-to-end optimization.

Matrix Operations: From Jones calculus to metric tensors:

y=A(θ)x,where ∂y∂θ=∂A∂θx(8)y = A(\theta)x, \qquad\text{where}\ \frac{\partial y}{\partial\theta} = \frac{\partial A}{\partial\theta}x \tag*{(8)}

Unexplored Applications for Automatic Differentiation

We now present nine astrophysical domains where, to our knowledge, automatic differentiation has not been applied for systematic exploration of solution spaces. For each application, we demonstrate how gradient-based discovery can reveal solutions beyond human intuition.

General Relativistic Applications

The Einstein field equations and related spacetime metrics constitute inherently differentiable systems, yet optimization approaches for exotic solutions remain largely unexplored.

Alcubierre Metric Optimization

The Alcubierre metric [11] represents a solution to the Einstein field equations permitting apparent faster-than-light travel through spacetime distortion:

ds2=−dt2+[dx−vsf(rs)dt]2+dy2+dz2(9)ds^2 = -dt^2 + [dx-v_s f(r_s)dt]^2 + dy^2 + dz^2 \tag*{(9)}

where f(rs)f(r_s) is the shape function and rs=[(x−xs(t))2+y2+z2]1/2r_s = [(x-x_s(t))^2+y^2+z^2]^{1/2}.

The stress-energy tensor components require:

T00=vs28π[(x−xs)2rs2(dfdrs)2+f2d2fdrs2](10)T^{00} = \frac{v_s^2}{8\pi}\left[\frac{(x-x_s)^2}{r_s^2}\left(\frac{df}{dr_s}\right)^2+f^2\frac{d^2f}{dr_s^2}\right] \tag*{(10)}

This expression assumes a static bubble configuration where x˙s=0\dot{x}_s=0. For a moving bubble, additional terms arise from the time dependence of rsr_s through xs(t)x_s(t). In our optimization framework, we consider static snapshots where the bubble position is fixed during each gradient computation, making the above expression valid.

Discovery Problem: Explore the landscape of all spacetime geometries that satisfy Einstein’s equations while minimizing ∫∣Tμν∣ dV\int|T^{\mu\nu}|\,dV, subject to boundary conditions f(0)=1f(0)=1, f(∞)=0f(\infty)=0.

Differentiable Formulation We parameterize the shape function using a neural network or smooth basis functions to ensure differentiability:

f(rs;θ)=∑k=1Kθkϕk(rs)(11)f(r_s;\theta)=\sum_{k=1}^{K}\theta_k\phi_k(r_s) \tag*{(11)}

where ϕk\phi_k are basis functions (e.g., radial basis functions, B-splines) and θ\theta are learnable parameters.

For neural parameterization:

f(rs;θ)=σ(WL⋯σ(W2σ(W1rs+b1)+b2)⋯+bL)(12)f(r_s;\theta)=\sigma\left(W_L\cdots\sigma\left(W_2\sigma\left(W_1r_s+b_1\right)+b_2\right)\cdots+b_L\right) \tag*{(12)}

where σ\sigma is a smooth activation function and θ={Wi,bi}\theta=\{W_i,b_i\}. Unlike traditional analytical forms that constrain the search to human-conceived geometries, neural parameterization enables the discovery of novel spacetime configurations that may violate geometric intuition while better satisfying the physical constraints.

Gradient Computation The key challenge is computing gradients through the stress-energy tensor. We decompose the computation:

  1. Shape function derivatives: For the basis function parameterization:

∂f∂rs=∑kθk∂ϕk∂rs,∂2f∂rs2=∑kθk∂2ϕk∂rs2(13)\frac{\partial f}{\partial r_s}=\sum_k\theta_k\frac{\partial\phi_k}{\partial r_s},\qquad\frac{\partial^2f}{\partial r_s^2}=\sum_k\theta_k\frac{\partial^2\phi_k}{\partial r_s^2} \tag*{(13)}

For neural network parameterization, we compute these derivatives directly through automatic differentiation of f(rs;θ)f(r_s;\theta).

  1. Stress-energy components: Each component is differentiable:

∂T00∂θ=vs28π[(x−xs)2rs2⋅2∂f∂rs∂∂θ(∂f∂rs)+2f∂f∂θ∂2f∂rs2+f2∂∂θ(∂2f∂rs2)](14)\frac{\partial T^{00}}{\partial\theta}=\frac{v_s^2}{8\pi}\left[\frac{(x-x_s)^2}{r_s^2}\cdot2\frac{\partial f}{\partial r_s}\frac{\partial}{\partial\theta}\left(\frac{\partial f}{\partial r_s}\right)+2f\frac{\partial f}{\partial\theta}\frac{\partial^2f}{\partial r_s^2}+f^2\frac{\partial}{\partial\theta}\left(\frac{\partial^2f}{\partial r_s^2}\right)\right] \tag*{(14)}
  1. Volume integral: We discretize on a spatial grid. For simplicity, we use Cartesian coordinates with uniform spacing:

E[θ]=∑i,j,k∣Tμν(xi,yj,zk;θ)∣ΔxΔyΔz(15)E[\theta]=\sum_{i,j,k}\left|T^{\mu\nu}(x_i,y_j,z_k;\theta)\right|\Delta x\Delta y\Delta z \tag*{(15)}

Note: For the logarithmic radial spacing mentioned in Eq. (16), appropriate volume weights would be needed.

Implementation Algorithm

Algorithm: Differentiable Alcubierre Optimization
Input: Initial parameters θ0\theta_0, velocity vsv_s, grid resolution NN
Output: Optimized shape function parameters θ∗\theta^*
1. Initialize spatial grid: {xi,yj,zk}\{x_i,y_j,z_k\} for i,j,k∈[1,N]i,j,k\in[1,N]
2. repeat
a. Compute rs(i,j,k)=xi2+yj2+zk2r_s(i,j,k)=\sqrt{x_i^2+y_j^2+z_k^2} for all grid points
b. Evaluate f(rs;θ)f(r_s;\theta) using neural network/basis functions
c. Compute derivatives ∂f∂rs\frac{\partial f}{\partial r_s}, ∂2f∂rs2\frac{\partial^2 f}{\partial r_s^2} via autodiff
d. Calculate TμνT^{\mu\nu} components at each grid point
e. Compute loss: L=1N3∑i,j,k∣Tμν(i,j,k)∣ΔxΔyΔz\mathcal{L}=\frac{1}{N^3}\sum_{i,j,k}|T^{\mu\nu}(i,j,k)|\Delta x\Delta y\Delta z
f. Add constraints: L←L+λ1(f(0)−1)2+λ2f(rmax⁡)2\mathcal{L}\leftarrow\mathcal{L}+\lambda_1(f(0)-1)^2+\lambda_2f(r_{\max})^2
g. Update: θ←θ−η∇θL\theta\leftarrow\theta-\eta\nabla_\theta\mathcal{L}
3. until convergence
Note: Weights λi\lambda_i should be tuned to balance terms (typically λ1,λ2∼102−104\lambda_1,\lambda_2\sim10^2-10^4 for normalized TμνT^{\mu\nu}).

Table 1.

Numerical Considerations

  • Spatial discretization: The grid must extend sufficiently far to capture the asymptotic behavior. We use logarithmic spacing near the origin:

ri=rmin⁡exp⁡(ilog⁡(rmax⁡/rmin⁡)Nr)(16)r_i = r_{\min} \exp\left(i\frac{\log(r_{\max}/r_{\min})}{N_r}\right) \tag*{(16)}
  • Boundary conditions: Enforce constraints using penalty terms:

Lboundary=λ1∣f(0)−1∣2+λ2∣f(rmax⁡)∣2+λ3∫0rmax⁡max⁡(0,−f(r))2dr(17)\mathcal{L}_{\mathrm{boundary}} = \lambda_1|f(0)-1|^2 + \lambda_2|f(r_{\max})|^2 + \lambda_3\int_0^{r_{\max}}\max(0,-f(r))^2\mathrm{d}r \tag*{(17)}
  • Gradient stability: The term (x−xs)2/rs2(x-x_s)^2/r_s^2 can cause numerical issues near rs=0r_s = 0. We regularize:

(x−xs)2rs2→(x−xs)2rs2+ϵ(18)\frac{(x-x_s)^2}{r_s^2} \rightarrow\frac{(x-x_s)^2}{r_s^2+\epsilon} \tag*{(18)}

with ϵ∼10−4\epsilon\sim10^{-4} for single-precision (float32) or ϵ∼10−8\epsilon\sim10^{-8} for double-precision (float64) computations.

  • Efficient implementation: Vectorize all operations over the spatial grid:

# Pseudocode
r_s = torch.sqrt(X**2 + Y**2 + Z**2 + eps)
f = shape_network(r_s)
df_dr = torch.autograd.grad(f, r_s, create_graph=True)[0]
d2f_dr2 = torch.autograd.grad(df_dr, r_s, create_graph=True)[0]
T_00 = compute_stress_energy(f, df_dr, d2f_dr2, v_s)
loss = torch.sum(torch.abs(T_00)) * dx * dy * dz

Geometry Space Exploration The neural parameterization transforms the problem from optimizing a single predetermined geometry to systematically exploring the entire landscape of physically valid spacetime configurations. We can implement several exploration strategies:

Multi-start optimization: Initialize multiple neural networks with different random weights to explore diverse regions of geometry space:

{θi∗}={arg⁡min⁡θiE[θi]},i=1,…,Nstarts(19)\{\theta_i^*\} = \{\arg\min_{\theta_i} E[\theta_i]\}, \quad i = 1,\ldots,N_{\mathrm{starts}} \tag*{(19)}

Geometry generation pipeline: Train a generative model to produce diverse initial geometries:

θinit∼pgenerator(θ∣constraints)(20)\theta_{\mathrm{init}} \sim p_{\mathrm{generator}}(\theta\mid\mathrm{constraints}) \tag*{(20)}

Population-based exploration: Combine gradient descent with evolutionary strategies to maintain diversity while optimizing:

θi(t+1)=θi(t)−η∇E[θi(t)]+ϵi(t)(21)\theta_i^{(t+1)} = \theta_i^{(t)} - \eta\nabla E[\theta_i^{(t)}] + \epsilon_i^{(t)} \tag*{(21)}

where ϵi(t)\epsilon_i^{(t)} introduces stochastic exploration.

This approach enables discovery of multiple solution families, characterization of the Pareto frontier between competing objectives (e.g., exotic matter vs. bubble size), and identification of qualitatively different geometric configurations that satisfy the same physical constraints. The framework transforms spacetime engineering from a problem of optimizing human-conceived designs to one of discovering the full spectrum of physically realizable geometries.

Cosmic String Network Evolution

Cosmic strings [12] evolve according to the Nambu-Goto action:

S=−μ∫d2σ−det⁡(γab),γab=∂aXμ∂bXμ(22)S = -\mu\int\mathrm{d}^2\sigma\sqrt{-\det(\gamma_{ab})}, \quad\gamma_{ab} = \partial_a X^\mu\partial_b X_\mu \tag*{(22)}

where Xμ(σ,τ)X^\mu(\sigma,\tau) describes the string worldsheet in spacetime coordinates μ=0,1,2,3\mu= 0,1,2,3 and σa=(σ,τ)\sigma^a = (\sigma,\tau) are worldsheet coordinates.

The equations of motion derived from (24) are:

∂a(−γγab∂bXμ)=0(23)\partial_a\left(\sqrt{-\gamma}\gamma^{ab}\partial_b X^\mu\right)=0 \tag*{(23)}

with γ=det⁡(γab)\gamma= \det(\gamma_{ab}) and γab\gamma^{ab} the inverse induced metric.

Discovery Problem: Given observational constraints on the stochastic gravitational wave background ΩGW(f)\Omega_{\mathrm{GW}}(f), find the string network parameters θ=(μ,preconnect,ξinitial)\theta= (\mu, p_{\mathrm{reconnect}}, \xi_{\mathrm{initial}}) that minimize:

L(θ)=∫∣ΩGWmodel(f;θ)−ΩGWobs(f)∣2 df+λR(θ),(24)\mathcal{L}(\theta) = \int\left|\Omega_{\mathrm{GW}}^{\mathrm{model}}(f;\theta) - \Omega_{\mathrm{GW}}^{\mathrm{obs}}(f)\right|^2 \, df + \lambda R(\theta), \tag*{(24)}

where R(θ)R(\theta) penalizes unphysical configurations (e.g., negative string tensions).

Differentiable Formulation:

  • Worldsheet Parameterization: Represent Xμ(σ,τ;θ)X^\mu(\sigma,\tau;\theta) as a spectral decomposition:

Xμ(σ,τ;θ)=∑k=1Kθkμϕk(σ,τ),(25)X^\mu(\sigma,\tau;\theta) = \sum_{k=1}^{K} \theta_k^\mu\phi_k(\sigma,\tau), \tag*{(25)}

where ϕk\phi_k are Fourier basis functions for periodic strings or B-splines for non-periodic segments.

  • Gradient Computation: The adjoint method computes gradients through the PDE constraint (23):

∂L∂θ=∫∂L∂Xμ∂Xμ∂θ d2σ+adjoint terms,(26)\frac{\partial\mathcal{L}}{\partial\theta} = \int\frac{\partial\mathcal{L}}{\partial X^\mu}\frac{\partial X^\mu}{\partial\theta} \, d^2\sigma+ \text{adjoint terms}, \tag*{(26)}

with the adjoint field λμ(σ,τ)\lambda^\mu(\sigma,\tau) satisfying:

∂a(−γγab∂bλμ)=−δLδXμ.(27)\partial_a\left(\sqrt{-\gamma}\gamma^{ab}\partial_b\lambda^\mu\right) = -\frac{\delta\mathcal{L}}{\delta X^\mu}. \tag*{(27)}

Algorithm: Differentiable Cosmic String Optimization

Input: Initial parameters θ0\theta_0, worldsheet resolution (Nσ,Nτ)(N_\sigma,N_\tau), observational data ΩGWobs(f)\Omega_{\mathrm{GW}}^{\mathrm{obs}}(f) Output: Optimized parameters θ∗\theta^*

  1. Initialize adaptive worldsheet grid (σi,τj)(\sigma_i,\tau_j)

  2. repeat a. Compute induced metric γab=∂aXμ∂bXμ\gamma_{ab} = \partial_a X^\mu\partial_b X_\mu b. Evaluate equation residual Rμ=∂a(−γγab∂bXμ)R^\mu= \partial_a\left(\sqrt{-\gamma}\gamma^{ab}\partial_b X^\mu\right) c. Calculate GW spectrum ΩGWmodel(f;θ)\Omega_{\mathrm{GW}}^{\mathrm{model}}(f;\theta) d. Compute loss L=∥Rμ∥2+α∥ΩGWmodel−ΩGWobs∥2\mathcal{L} = \lVert R^\mu\rVert^2 + \alpha\lVert\Omega_{\mathrm{GW}}^{\mathrm{model}} - \Omega_{\mathrm{GW}}^{\mathrm{obs}}\rVert^2 e. Solve adjoint equation for λμ(σ,τ)\lambda^\mu(\sigma,\tau) f. Update θ←θ−η∇θL\theta\leftarrow\theta- \eta\nabla_\theta\mathcal{L}

  3. until convergence

Numerical Considerations:

  • Singularity Handling: The worldsheet metric γab\gamma_{ab} becomes degenerate at cusps (det⁡(γ)→0\det(\gamma) \to0). We regularize with:

−γ→−γ+ϵ,ϵ∼10−8.(28)\sqrt{-\gamma} \to\sqrt{-\gamma+ \epsilon}, \qquad\epsilon\sim10^{-8}. \tag*{(28)}
  • Discretization: Use adaptive mesh refinement near high-curvature regions, with gradient-aware refinement criteria:

Δσ∝∥∂L∂Xμ∥−1/2.(29)\Delta\sigma\propto\left\lVert\frac{\partial\mathcal{L}}{\partial X^\mu}\right\rVert^{-1/2}. \tag*{(29)}

Exploration Strategies:

  • Multi-Objective Optimization: Simultaneously minimize L(θ)\mathcal{L}(\theta) for multiple GW datasets:

    • Pulsar Timing Arrays (PTA) for nanohertz signals

    • LIGO/Virgo/KAGRA for high-frequency bursts

  • Physics-Informed Initialization: Seed θ\theta with values from analytic string network models [20].

Domain Wall Dynamics

Domain walls [12] are topological defects described by the scalar field Lagrangian:

L=12∂μϕ∂μϕ−V(ϕ),V(ϕ)=λ4(ϕ2−η2)2.(30)\mathcal{L} = \frac{1}{2}\partial_\mu\phi\partial^\mu\phi- V(\phi), \quad V(\phi) = \frac{\lambda}{4}(\phi^2-\eta^2)^2. \tag*{(30)}

For a static planar wall at z=0z=0, the field equation reduces to:

d2ϕdz2=λϕ(ϕ2−η2),ϕ(−∞)=−η, ϕ(+∞)=+η.(31)\frac{d^2\phi}{dz^2} = \lambda\phi(\phi^2-\eta^2), \quad\phi(-\infty)=-\eta,\ \phi(+\infty)=+\eta. \tag*{(31)}

Discovery Problem: Find field configurations ϕ(z;θ)\phi(z;\theta) that minimize the total energy

E[θ]=∫−∞∞[12(dϕdz)2+V(ϕ)] dz,(32)E[\theta] = \int_{-\infty}^{\infty}\left[\frac{1}{2}\left(\frac{d\phi}{dz}\right)^2 + V(\phi)\right]\,dz, \tag*{(32)}

while satisfying observational constraints from CMB anisotropies [21].

Differentiable Formulation Parameterize ϕ(z;θ)\phi(z;\theta) using a spectral decomposition:

ϕ(z;θ)=ηtanh⁡(zδ0)+∑k=1Kθkψk(z),(33)\phi(z;\theta) = \eta\tanh\left(\frac{z}{\delta_0}\right) + \sum_{k=1}^{K}\theta_k\psi_k(z), \tag*{(33)}

where δ0=2/λη−1\delta_0=\sqrt{2/\lambda\eta^{-1}} and ψk(z)\psi_k(z) are compactly supported basis functions (e.g., B-splines).

The adjoint field ξ(z)\xi(z) for gradient computation satisfies:

d2ξdz2−λ(3ϕ2−η2)ξ=δEδϕ,ξ(±∞)=0.(34)\frac{d^2\xi}{dz^2} - \lambda(3\phi^2-\eta^2)\xi= \frac{\delta E}{\delta\phi}, \quad\xi(\pm\infty)=0. \tag*{(34)}

Numerical Considerations

  • Grid Construction: Use coordinate mapping z→z~=z/(1+∣z∣)z\to\tilde{z}=z/(1+|z|) to compactify the domain to [−1,1][-1,1].

  • Regularization: Add a small kinetic term to prevent singularities:

E→E+ϵ∫(d2ϕdz2)2 dz,ϵ∼10−4.(35)E\to E+\epsilon\int\left(\frac{d^2\phi}{dz^2}\right)^2\,dz,\quad\epsilon\sim10^{-4}. \tag*{(35)}

Exploration Strategies

  • Multi-Objective Optimization: Simultaneously minimize:

L1=E[θ],L2=∥Cℓmodel(θ)−Cℓobs∥2.(36)\mathcal{L}_1=E[\theta],\quad\mathcal{L}_2=\left\|C_\ell^{\mathrm{model}}(\theta)-C_\ell^{\mathrm{obs}}\right\|^2. \tag*{(36)}
  • Analytic Baseline: Compare against the exact tanh solution for verification.

Domain walls [12] are topological defects described by the scalar field Lagrangian:

L=12∂μϕ∂μϕ−V(ϕ),V(ϕ)=λ4(ϕ2−η2)2.(37)\mathcal{L} = \frac{1}{2}\partial_\mu\phi\partial^\mu\phi- V(\phi), \quad V(\phi) = \frac{\lambda}{4}(\phi^2-\eta^2)^2. \tag*{(37)}

For a static planar wall at z=0z=0, the field equation reduces to:

d2ϕdz2=λϕ(ϕ2−η2),ϕ(−∞)=−η, ϕ(+∞)=+η.(38)\frac{d^2\phi}{dz^2} = \lambda\phi(\phi^2-\eta^2), \quad\phi(-\infty)=-\eta,\ \phi(+\infty)=+\eta. \tag*{(38)}

Discovery Problem: Find field configurations ϕ(z;θ)\phi(z;\theta) that minimize the total energy

E[θ]=∫−∞∞[12(dϕdz)2+V(ϕ)] dz,(39)E[\theta] = \int_{-\infty}^{\infty}\left[\frac{1}{2}\left(\frac{d\phi}{dz}\right)^2 + V(\phi)\right]\,dz, \tag*{(39)}

while satisfying observational constraints from CMB anisotropies [21].

Differentiable Formulation Parameterize ϕ(z;θ)\phi(z;\theta) using a spectral decomposition:

ϕ(z;θ)=ηtanh⁡(zδ0)+∑k=1Kθkψk(z),(40)\phi(z;\theta)=\eta\tanh\left(\frac{z}{\delta_0}\right)+\sum_{k=1}^{K}\theta_k\psi_k(z), \tag*{(40)}

where δ0=2/λη−1\delta_0=\sqrt{2/\lambda\eta^{-1}} and ψk(z)\psi_k(z) are compactly supported basis functions (e.g., B-splines).

The adjoint field ξ(z)\xi(z) for gradient computation satisfies:

d2ξdz2−λ(3ϕ2−η2)ξ=δEδϕ,ξ(±∞)=0.(41)\frac{d^2\xi}{dz^2}-\lambda(3\phi^2-\eta^2)\xi=\frac{\delta E}{\delta\phi},\qquad\xi(\pm\infty)=0. \tag*{(41)}

Numerical Considerations

  • Grid Construction: Use coordinate mapping z→z~=z/(1+∣z∣)z\to\tilde{z}=z/(1+|z|) to compactify the domain to [−1,1][-1,1].

  • Regularization: Add a small kinetic term to prevent singularities:

E→E+ϵ∫(d2ϕdz2)2dz,ϵ∼10−4.(42)E\to E+\epsilon\int\left(\frac{d^2\phi}{dz^2}\right)^2 dz,\qquad\epsilon\sim10^{-4}. \tag*{(42)}

Exploration Strategies

  • Multi-Objective Optimization: Simultaneously minimize:

L1=E[θ],L2=∥Cℓmodel(θ)−Cℓobs∥2.(43)\mathcal{L}_1=E[\theta],\qquad\mathcal{L}_2=\lVert C_\ell^{\mathrm{model}}(\theta)-C_\ell^{\mathrm{obs}}\rVert^2. \tag*{(43)}
  • Analytic Baseline: Compare against the exact tanh solution for verification.

False Vacuum Decay and Bubble Nucleation

The nucleation rate of bubbles in a metastable vacuum [14] is determined by the bounce solution minimizing the Euclidean action:

SE=2π2∫0∞dr r3[12(dϕdr)2+V(ϕ)](44)S_E=2\pi^2\int_0^\infty dr\,r^3\left[\frac{1}{2}\left(\frac{d\phi}{dr}\right)^2+V(\phi)\right] \tag*{(44)}

The bounce equation derived from δSE=0\delta S_E=0 is:

d2ϕdr2+3rdϕdr=dVdϕ(45)\frac{d^2\phi}{dr^2}+\frac{3}{r}\frac{d\phi}{dr}=\frac{dV}{d\phi} \tag*{(45)}

with boundary conditions ϕ(0)=ϕtop\phi(0)=\phi_{\mathrm{top}} (field value at the potential barrier) and ϕ(∞)=ϕfalse\phi(\infty)=\phi_{\mathrm{false}} (false vacuum state).

Discovery Problem: Find all field configurations ϕ(r)\phi(r) that minimize SES_E while satisfying quantum tunneling constraints, determining the decay rate Γ∼e−SE\Gamma\sim e^{-S_E}.

Differentiable Formulation We parameterize the bounce solution using a neural network with radial symmetry:

ϕ(r;θ)=ϕfalse+(ϕtop−ϕfalse)σ(WL[…σ(W1r+b1)…]+bL)(46)\phi(r;\theta)=\phi_{\mathrm{false}}+(\phi_{\mathrm{top}}-\phi_{\mathrm{false}})\sigma(W_L[\ldots\sigma(W_1r+b_1)\ldots]+b_L) \tag*{(46)}

where σ\sigma is a smooth activation function (tanh) and θ={Wi,bi}\theta=\{W_i,b_i\} are learnable parameters. This automatically satisfies ϕ(0)≈ϕtop\phi(0)\approx\phi_{\mathrm{top}} through initialization and ϕ(∞)=ϕfalse\phi(\infty)=\phi_{\mathrm{false}} via activation function asymptotics.

Gradient Computation The AD workflow requires:

  1. Forward solution of the ODE (Eq. (2)) using the parameterized ϕ(r;θ)\phi(r;\theta)

  2. Evaluation of SE[θ]S_E[\theta] via adaptive quadrature

  3. Backpropagation through the ODE solver using adjoint methods

  4. Constraint enforcement via penalty terms:

L=SE[θ]+λ1∣ϕ(0)−ϕtop∣2+λ2∣ϕ(Rmax⁡)−ϕfalse∣2(47)\mathcal{L}=S_E[\theta]+\lambda_1|\phi(0)-\phi_{\mathrm{top}}|^2+\lambda_2|\phi(R_{\max})-\phi_{\mathrm{false}}|^2 \tag*{(47)}

Algorithm: Differentiable Bubble Optimization

Input: Initial parameters θ0\theta_0, potential V(ϕ)V(\phi), grid {ri}i=1N\{r_i\}_{i=1}^{N}

Output: Optimized bounce solution ϕ∗(r)\phi^*(r)

  1. Initialize radial grid with logarithmic spacing ri=rmin⁡exp⁡(iΔ)r_i = r_{\min}\exp(i\Delta)

  2. repeat a. Compute ϕ(ri;θ)\phi(r_i;\theta) and derivatives via autodiff b. Evaluate ODE residual R=ϕ′′+3rϕ′−V′(ϕ)R = \phi^{\prime\prime} + \frac{3}{r}\phi^\prime- V^\prime(\phi) c. Calculate loss L=SE[θ]+λ∥R∥2\mathcal{L} = S_E[\theta] + \lambda\lVert R\rVert^2 d. Update θ←θ−η∇θL\theta\leftarrow\theta- \eta\nabla_\theta\mathcal{L}

  3. until convergence (∥∇L∥<ϵ\lVert\nabla\mathcal{L}\rVert< \epsilon)

Note: λ\lambda should be progressively increased during optimization.

Numerical Considerations

  • Singularity Handling: Regularize the 1/r1/r term:

3r→3r+ϵ,ϵ∼10−8(48)\frac{3}{r} \rightarrow\frac{3}{r+\epsilon}, \qquad\epsilon\sim10^{-8} \tag*{(48)}
  • Adaptive Mesh: Refine grid where ∣dV/dϕ∣|dV/d\phi| is large:

Δrnew=C∣V′(ϕ)∣+δ(49)\Delta r_{\mathrm{new}} = \frac{C}{|V^\prime(\phi)|+\delta} \tag*{(49)}
  • Potential Smoothing: Apply Gaussian filtering if V(ϕ)V(\phi) has discontinuities:

Vsmooth=V∗N(0,σ2)(50)V_{\mathrm{smooth}} = V * \mathcal{N}(0,\sigma^2) \tag*{(50)}

Solution Space Exploration

  • Multi-Field Tunneling: Extend to ϕ⃗=(ϕ1,…,ϕN)\vec{\phi} = (\phi_1,\ldots,\phi_N) with:

SE=2π2∫r3[12∑i(dϕidr)2+V(ϕ⃗)] dr(51)S_E = 2\pi^2 \int r^3 \left[\frac{1}{2}\sum_i\left(\frac{d\phi_i}{dr}\right)^2 + V(\vec{\phi})\right]\,dr \tag*{(51)}
  • Temperature Effects: Include finite-TT via compact time dimension:

SET=SE+∫0βdτ∫d3x[12(∂τϕ)2+V(ϕ)](52)S_E^T = S_E + \int_0^\beta d\tau\int d^3x \left[\frac{1}{2}(\partial_\tau\phi)^2 + V(\phi)\right] \tag*{(52)}
  • Non-Canonical Kinetics: Generalize for field-space metrics Gij(ϕ)G_{ij}(\phi):

SE=2π2∫r3[12Gij(ϕ)dϕidrdϕjdr+V(ϕ)] dr(53)S_E = 2\pi^2 \int r^3 \left[\frac{1}{2}G_{ij}(\phi)\frac{d\phi_i}{dr}\frac{d\phi_j}{dr} + V(\phi)\right]\,dr \tag*{(53)}

Radio Astronomy and Interferometry

Radio interferometry fundamentally involves complex-valued measurements and calibration challenges that are naturally suited to gradient-based optimization, yet current methods rely on alternating least squares or stochastic sampling without exploiting gradient information.

Interferometric Calibration

Radio interferometers measure complex visibilities that encode spatial Fourier components of the sky brightness distribution. The measurement process introduces complex gains that must be calibrated to recover accurate astronomical information [15].

The measurement equation for baseline ijij is:

Vijobs(t,ν)=gi(t,ν)gj∗(t,ν)Vijtrue(t,ν)+nij(t,ν)(54)V_{ij}^{\mathrm{obs}}(t,\nu)=g_i(t,\nu)g_j^*(t,\nu)V_{ij}^{\mathrm{true}}(t,\nu)+n_{ij}(t,\nu) \tag*{(54)}

where gig_i are complex antenna gains and nijn_{ij} is thermal noise.

For direction-dependent effects (ionosphere, primary beam), the equation becomes:

Vijobs=∑k=1NdirJikVijkJjk†+nij(55)V_{ij}^{\mathrm{obs}}=\sum_{k=1}^{N_{\mathrm{dir}}}J_{i}^{k}V_{ij}^{k}J_{j}^{k\dagger}+n_{ij} \tag*{(55)}

where JikJ_i^k are 2×22 \times2 Jones matrices and VijkV_{ij}^{k} are visibility coherency matrices.

Discovery Problem: Explore the space of calibration solutions that minimize:

L(θ)=∑ij,t,ν∣Vijobs−M(Vijmodel;θ)∣2σij2+R(θ)(56)\mathcal{L}(\theta)=\sum_{ij,t,\nu}\frac{\left|V_{ij}^{\mathrm{obs}}-\mathcal{M}(V_{ij}^{\mathrm{model}};\theta)\right|^2}{\sigma_{ij}^2}+\mathcal{R}(\theta) \tag*{(56)}

where θ={gi,ϕion,Jik}\theta=\{g_i,\phi_{\mathrm{ion}},J_i^k\} includes all calibration parameters and R\mathcal{R} enforces physical constraints.

Differentiable Formulation We parameterize the complex gains and ionospheric phases using smooth functions:

gi(t,ν)=ai(t,ν)exp⁡(iϕi(t,ν))(57)g_i(t,\nu)=a_i(t,\nu)\exp(i\phi_i(t,\nu)) \tag*{(57)}

where amplitudes and phases are parameterized as:

ai(t,ν)=exp⁡(∑k=0Kaθa,kiBk(t)Pk(ν))(58)a_i(t,\nu)=\exp\left(\sum_{k=0}^{K_a}\theta_{a,k}^{i}B_k(t)P_k(\nu)\right) \tag*{(58)}
ϕi(t,ν)=∑k=0Kϕθϕ,kiBk(t)Pk(ν)+ϕioni(t)(59)\phi_i(t,\nu)=\sum_{k=0}^{K_\phi}\theta_{\phi,k}^{i}B_k(t)P_k(\nu)+\phi_{\mathrm{ion}}^{i}(t) \tag*{(59)}

with BkB_k being B-spline basis functions in time and PkP_k polynomials in frequency.

The ionospheric contribution is modeled as a 2D phase screen:

ϕioni(t)=8.45×109ν∫LOSine(x,y,z,t) dl(60)\phi_{\mathrm{ion}}^{i}(t)=\frac{8.45\times10^{9}}{\nu}\int_{\mathrm{LOS}_i}n_e(x,y,z,t)\,dl \tag*{(60)}

We parameterize the electron density using a neural network:

ne(x,y,z,t;θion)=NNθion(x,y,z,t)(61)n_e(x,y,z,t;\theta_{\mathrm{ion}})=\mathrm{NN}_{\theta_{\mathrm{ion}}}(x,y,z,t) \tag*{(61)}

Gradient Computation The complex-valued nature requires Wirtinger derivatives:

∂L∂gi=∑jgj∗Vijmodel(Vijobs−gigj∗Vijmodel)∗σij2(62)\frac{\partial\mathcal{L}}{\partial g_i}=\sum_j\frac{g_j^*V_{ij}^{\mathrm{model}}\left(V_{ij}^{\mathrm{obs}}-g_ig_j^*V_{ij}^{\mathrm{model}}\right)^*}{\sigma_{ij}^2} \tag*{(62)}
∂L∂gi∗=∑jgjVijmodel∗(Vijobs−gigj∗Vijmodel)σij2(63)\frac{\partial\mathcal{L}}{\partial g_i^*}=\sum_j\frac{g_jV_{ij}^{\mathrm{model}*}\left(V_{ij}^{\mathrm{obs}}-g_ig_j^*V_{ij}^{\mathrm{model}}\right)}{\sigma_{ij}^2} \tag*{(63)}

For the ionospheric screen, we use the chain rule:

∂L∂θion=∑i∂L∂ϕioni∂ϕioni∂ne∂ne∂θion(64)\frac{\partial\mathcal{L}}{\partial\theta_{\mathrm{ion}}}=\sum_i\frac{\partial\mathcal{L}}{\partial\phi_{\mathrm{ion}}^{i}}\frac{\partial\phi_{\mathrm{ion}}^{i}}{\partial n_e}\frac{\partial n_e}{\partial\theta_{\mathrm{ion}}} \tag*{(64)}

Algorithm: Differentiable Interferometric Calibration

Input: Observed visibilities VijobsV_{ij}^{\mathrm{obs}}, sky model VijmodelV_{ij}^{\mathrm{model}}, initial θ0\theta_0

Output: Optimized calibration parameters θ∗\theta^*

  1. Initialize gain amplitudes ai=1a_i = 1, phases ϕi=0\phi_i = 0

  2. Precompute basis functions Bk(t)B_k(t), Pk(ν)P_k(\nu) on measurement grid

  3. repeat a. Compute complex gains gi=aiexp⁡(iϕi)g_i = a_i \exp(i\phi_i) b. Apply gains: Vijcal=gigj∗VijmodelV_{ij}^{\mathrm{cal}} = g_i g_j^* V_{ij}^{\mathrm{model}} c. Calculate residuals Rij=Vijobs−VijcalR_{ij} = V_{ij}^{\mathrm{obs}} - V_{ij}^{\mathrm{cal}} d. Compute loss L=∑ij∣Rij∣2/σij2\mathcal{L} = \sum_{ij} \lvert R_{ij}\rvert^2/\sigma_{ij}^2 e. Add regularization:

    • Smoothness: L←L+λ1∑i∥∇tgi∥2\mathcal{L} \leftarrow\mathcal{L} + \lambda_1 \sum_i \lVert\nabla_t g_i\rVert^2

    • Unity constraint: L←L+λ2∑i(∣gi∣−1)2\mathcal{L} \leftarrow\mathcal{L} + \lambda_2 \sum_i (\lvert g_i\rvert- 1)^2 f. Compute Wirtinger derivatives ∂L/∂gi\partial\mathcal{L}/\partial g_i, ∂L/∂gi∗\partial\mathcal{L}/\partial g_i^* g. Update θ←θ−η∇θL\theta\leftarrow\theta- \eta\nabla_\theta\mathcal{L}

  4. until ∥∇L∥<ϵ\lVert\nabla\mathcal{L}\rVert< \epsilon or max iterations

Numerical Considerations

  • Gauge Degeneracy: The product gigj∗g_i g_j^* is invariant under gi→gieiαg_i \rightarrow g_i e^{i\alpha}. Fix by constraining one reference antenna:

gref=∣gref∣,Im⁡(gref)=0(65)g_{\mathrm{ref}} = \lvert g_{\mathrm{ref}}\rvert,\quad\operatorname{Im}(g_{\mathrm{ref}}) = 0 \tag*{(65)}
  • Frequency Dependence: For wide bandwidths, model frequency-dependent gains:

gi(ν)=gi0(νν0)αiexp⁡(iτi(ν−ν0))(66)g_i(\nu) = g_i^0 \left(\frac{\nu}{\nu_0}\right)^{\alpha_i}\exp\left(i\tau_i(\nu-\nu_0)\right) \tag*{(66)}
  • Outlier Robustness: Replace L2L_2 loss with Huber loss:

ρ(x)={12x2∣x∣≤δδ(∣x∣−12δ)∣x∣>δ(67)\rho(x) = \begin{cases} \frac{1}{2}x^2 & \lvert x\rvert\le\delta\\ \delta(\lvert x\rvert- \frac{1}{2}\delta) & \lvert x\rvert> \delta \end{cases} \tag*{(67)}

Solution Space Exploration

  • Multi-Scale Optimization: Hierarchical approach from coarse to fine time/frequency resolution:

θ(k+1)=arg min⁡θ L(θ;Δtk,Δνk),Δtk+1=Δtk/2(68)\theta^{(k+1)} = \underset{\theta}{\operatorname{arg\,min}}\ \mathcal{L}(\theta;\Delta t_k,\Delta\nu_k),\quad\Delta t_{k+1} = \Delta t_k/2 \tag*{(68)}
  • Direction-Dependent Calibration: Parameterize Jones matrices using quaternions for SU(2)SU(2) constraint:

J=q0I+i(q1σx+q2σy+q3σz),∑iqi2=1(69)\mathbf{J} = q_0\mathbf{I} + i(q_1\sigma_x + q_2\sigma_y + q_3\sigma_z),\quad\sum_i q_i^2 = 1 \tag*{(69)}
  • Joint Sky-Calibration: Simultaneously optimize sky model and calibration:

min⁡Isky,θcal ∑ij∣Vijobs−gigj∗F(Isky)∣2σij2+λTV(Isky)(70)\underset{I_{\mathrm{sky}},\theta_{\mathrm{cal}}}{\operatorname{min}}\ \sum_{ij}\frac{\left\lvert V_{ij}^{\mathrm{obs}} - g_i g_j^* \mathcal{F}(I_{\mathrm{sky}})\right\rvert^2}{\sigma_{ij}^2} + \lambda TV(I_{\mathrm{sky}}) \tag*{(70)}

Black Hole Image Reconstruction

Very Long Baseline Interferometry (VLBI) observations of black holes [16] present an extreme sparse sampling problem in the Fourier domain. The Event Horizon Telescope’s success relied on regularized maximum likelihood methods that could benefit from gradient-based exploration of the image and hyperparameter space.

The forward model relates the image I(x,y)I(x,y) to observed visibilities:

V(u,v)=∬I(x,y)exp⁡[−2πi(ux+vy)] dxdy(71)V(u,v)=\iint I(x,y)\exp[-2\pi i(ux+vy)]\,\mathrm{d}x\mathrm{d}y \tag*{(71)}

where (u,v)(u,v) are spatial frequencies sampled by the interferometer.

For black hole imaging, relativistic effects modify the relationship:

Iobs(α,β)=∫γjνg3e−τν dλ(72)I_{\mathrm{obs}}(\alpha,\beta)=\int_{\gamma}j_{\nu}g^{3}e^{-\tau_{\nu}}\,\mathrm{d}\lambda \tag*{(72)}

where the integral follows null geodesics γ\gamma in the Kerr metric, jνj_{\nu} is the emission coefficient, gg is the gravitational redshift factor, and τν\tau_{\nu} is the optical depth.

Discovery Problem: Explore the space of physically valid black hole images that satisfy:

L(I,λ)=χdata2(I)+∑kλkRk(I)(73)\mathcal{L}(I,\lambda)=\chi_{\mathrm{data}}^{2}(I)+\sum_{k}\lambda_{k}\mathcal{R}_{k}(I) \tag*{(73)}

where Rk\mathcal{R}_{k} are regularizers encoding physical priors.

Differentiable Formulation We parameterize the image using a continuous representation:

I(x,y;θ)=∑k=1Kθkϕk(x,y)+INN(x,y;θNN)(74)I(x,y;\theta)=\sum_{k=1}^{K}\theta_{k}\phi_{k}(x,y)+I_{NN}(x,y;\theta_{NN}) \tag*{(74)}

where ϕk\phi_{k} are fixed basis functions (e.g., shapelets, wavelets) and INNI_{NN} is a neural network component for capturing fine structure.

The data fidelity term with closure quantities:

χdata2=χvis2+χamp2+χcphase2+χlogcamp2(75)\chi_{\mathrm{data}}^{2}=\chi_{\mathrm{vis}}^{2}+\chi_{\mathrm{amp}}^{2}+\chi_{\mathrm{cphase}}^{2}+\chi_{\mathrm{logcamp}}^{2} \tag*{(75)}

For closure phases (which cancel station-based errors):

χcphase2=∑ijk∣arg⁡(VijVjkVki)−arg⁡(V~ijV~jkV~ki)∣2σcphase2(76)\chi_{\mathrm{cphase}}^{2}=\sum_{ijk}\frac{\left|\arg(V_{ij}V_{jk}V_{ki})-\arg(\widetilde{V}_{ij}\widetilde{V}_{jk}\widetilde{V}_{ki})\right|^{2}}{\sigma_{\mathrm{cphase}}^{2}} \tag*{(76)}

Gradient Computation The key challenge is differentiating through the Fourier transform and closure quantities:

  1. Fourier Transform: Using the FFT with appropriate padding:

∂L∂I=F−1[∂L∂V~](77)\frac{\partial\mathcal{L}}{\partial I}=\mathcal{F}^{-1}\left[\frac{\partial\mathcal{L}}{\partial\widetilde{V}}\right] \tag*{(77)}
  1. Closure Phases: Require careful handling of phase wrapping:

∂χcphase2∂Vij=∑k2(Δϕijk)σijk2∂arg⁡(VijVjkVki)∂Vij(78)\frac{\partial\chi_{\mathrm{cphase}}^{2}}{\partial V_{ij}}=\sum_{k}\frac{2(\Delta\phi_{ijk})}{\sigma_{ijk}^{2}}\frac{\partial\arg(V_{ij}V_{jk}V_{ki})}{\partial V_{ij}} \tag*{(78)}

where the argument derivative uses:

∂arg⁡(z)∂z=12i(1z−1z∗)(79)\frac{\partial\arg(z)}{\partial z}=\frac{1}{2i}\left(\frac{1}{z}-\frac{1}{z^{*}}\right) \tag*{(79)}
  1. Ray Tracing: For Kerr metric ray tracing, use automatic differentiation through the geodesic equations:

d2xμdλ2+Γαβμdxαdλdxβdλ=0(80)\frac{\mathrm{d}^{2}x^{\mu}}{\mathrm{d}\lambda^{2}}+\Gamma^{\mu}_{\alpha\beta}\frac{\mathrm{d}x^{\alpha}}{\mathrm{d}\lambda}\frac{\mathrm{d}x^{\beta}}{\mathrm{d}\lambda}=0 \tag*{(80)}

Algorithm: Differentiable Black Hole Imaging

Input: Visibility data {Vijobs}\{V^{\mathrm{obs}}_{ij}\}, initial image I0I_0, regularization weights {λk}\{\lambda_k\}

Output: Reconstructed image I∗I^* and optimized hyperparameters λ∗\lambda^*

  1. Initialize image with Gaussian blob: I0(x,y)=Aexp⁡(−(x2+y2)/2σ2)I_0(x,y)=A\exp(-(x^2+y^2)/2\sigma^2)

  2. Precompute (u,v)(u,v) coverage and noise statistics

  3. repeat a. Compute model visibilities: V~=F(I)\widetilde{V}=\mathcal{F}(I) b. Interpolate to observed (u,v)(u,v) points using differentiable kernel c. Calculate closure quantities from model d. Compute data terms:

    • χvis2=∑ij∣Vijobs−V~ij∣2/σij2\chi^2_{\mathrm{vis}}=\sum_{ij}|V^{\mathrm{obs}}_{ij}-\widetilde{V}_{ij}|^2/\sigma_{ij}^2

    • χcphase2=∑ijk(ϕijkobs−ϕ~ijk)2/σϕ2\chi^2_{\mathrm{cphase}}=\sum_{ijk}(\phi^{\mathrm{obs}}_{ijk}-\widetilde{\phi}_{ijk})^2/\sigma_\phi^2

    e. Add regularization terms:

    • Total variation: TV(I)=∑x,y(∇xI)2+(∇yI)2+ϵTV(I)=\sum_{x,y}\sqrt{(\nabla_x I)^2+(\nabla_y I)^2+\epsilon}

    • Entropy: S(I)=−∑x,yIlog⁡(I/Iprior)S(I)=-\sum_{x,y}I\log(I/I_{\mathrm{prior}})

    • Ring prior: R(I)=∑x,yI(x,y)Kring(x,y)R(I)=\sum_{x,y}I(x,y)K_{\mathrm{ring}}(x,y)

    f. Update image: I←I−ηI∇ILI\leftarrow I-\eta_I\nabla_I\mathcal{L}

    g. Update hyperparameters: λ←λ−ηλ∇λL\lambda\leftarrow\lambda-\eta_\lambda\nabla_\lambda\mathcal{L}

  4. until convergence

Numerical Considerations

  • Positivity Constraint: Enforce I≥0I\geq0 using projected gradient descent:

Ik+1=max⁡(0,Ik−η∇L)(81)I_{k+1}=\max(0,I_k-\eta\nabla\mathcal{L}) \tag*{(81)}

or parameterize I=exp⁡(I~)I=\exp(\widetilde{I}) to ensure positivity.

  • Multi-Resolution: Use wavelet decomposition for efficient large-scale optimization:

I(x,y)=∑j,k,lcjklψjkl(x,y)(82)I(x,y)=\sum_{j,k,l}c_{jkl}\psi_{jkl}(x,y) \tag*{(82)}

where ψjkl\psi_{jkl} are wavelet basis functions at scale jj.

  • Fourier Interpolation: Use Kaiser-Bessel kernels for accurate gradients:

V~(u,v)=∑m,nIm,nK(u−um,v−vn)(83)\widetilde{V}(u,v)=\sum_{m,n}I_{m,n}K(u-u_m,v-v_n) \tag*{(83)}

Exploration Strategies

  • Multi-Frequency Synthesis: Joint reconstruction across observing bands:

L=∑νwνχdata2(Iν)+λspec∑x,y∥∇νI(x,y,ν)∥2(84)\mathcal{L}=\sum_{\nu}w_\nu\chi^2_{\mathrm{data}}(I_\nu)+\lambda_{\mathrm{spec}}\sum_{x,y}\lVert\nabla_\nu I(x,y,\nu)\rVert^2 \tag*{(84)}
  • Polarimetric Imaging: Extend to full Stokes parameters:

Vij=(VI+VQVU+iVVVU−iVVVI−VQ)ij(85)V_{ij}= \begin{pmatrix} V_I+V_Q & V_U+iV_V \\ V_U-iV_V & V_I-V_Q \end{pmatrix}_{ij} \tag*{(85)}
  • Dynamic Imaging: For time-variable sources:

min⁡I(x,y,t)∑tχdata2(It)+λtemp∫∣∂tI∣2 dt(86)\min_{I(x,y,t)}\sum_t\chi^2_{\mathrm{data}}(I_t)+\lambda_{\mathrm{temp}}\int|\partial_t I|^2\,dt \tag*{(86)}
  • Bayesian Hyperparameter Selection: Optimize marginal likelihood:

λ∗=arg⁡max⁡λ log⁡p(Vobs∣λ)=arg⁡max⁡λ[−12vTC−1v−12log⁡∣C∣](87)\lambda^* = \underset{\lambda}{\arg\max}\ \log p(V^{obs}\mid\lambda) = \underset{\lambda}{\arg\max}\left[-\frac{1}{2}v^T C^{-1}v-\frac{1}{2}\log\lvert C\rvert\right] \tag*{(87)}

where C=Cdata+Cprior(λ)C = C_{data} + C_{prior}(\lambda).

The differentiable formulation enables systematic exploration of the image space, discovering features that traditional methods might miss due to local minima, while simultaneously optimizing regularization parameters that balance data fidelity against physical priors. This transforms black hole imaging from an art requiring manual hyperparameter tuning to a principled discovery process.

Stellar and Planetary Astrophysics

Stellar and planetary systems present optimization problems spanning vast scales—from stellar cores where nuclear fusion occurs to planetary surfaces where thermal forces alter orbits. Despite smooth parameter dependencies and well-posed forward models, these domains remain largely unexplored by gradient-based methods.

Asteroseismic Inversion

Stars pulsate in global oscillation modes that probe their internal structure [17]. These acoustic and gravity waves satisfy the stellar pulsation equations, with frequencies encoding information about density stratification, sound speed profiles, and rotation. Current inversion methods rely on grid searches or genetic algorithms without exploiting gradient information.

The stellar oscillation equations in the adiabatic approximation are:

dξrdr=(2r−1Γ1dln⁡pdr)ξr+1Γ1p(Sℓ2ω2−1)P′(88)\frac{d\xi_r}{dr} = \left(\frac{2}{r}-\frac{1}{\Gamma_1}\frac{d\ln p}{dr}\right)\xi_r+\frac{1}{\Gamma_1p}\left(\frac{S_\ell^2}{\omega^2}-1\right)P' \tag*{(88)}
dP′dr=ρ(ω2−N2)ξr+dln⁡pdrP′(89)\frac{dP'}{dr} = \rho(\omega^2-N^2)\xi_r+\frac{d\ln p}{dr}P' \tag*{(89)}

where ξr\xi_r is the radial displacement, P′P' is the Eulerian pressure perturbation, SℓS_\ell is the Lamb frequency, and NN is the Brunt-Väisälä frequency.

The eigenvalue problem Lξ=ω2ξ\mathcal{L}\xi=\omega^2\xi yields discrete frequencies {ωnℓ}\{\omega_{n\ell}\} that depend on stellar structure parameters θ={M,R,Y0,Z,αMLT,fov,…}\theta=\{M,R,Y_0,Z,\alpha_{MLT},f_{ov},\ldots\}.

Discovery Problem: Given observed frequencies νnℓobs\nu_{n\ell}^{obs} with uncertainties σnℓ\sigma_{n\ell}, explore the space of stellar models that minimize:

L(θ)=∑n,ℓ(νnℓobs−νnℓmodel(θ))2σnℓ2+∑kλkRk(θ)(90)\mathcal{L}(\theta)=\sum_{n,\ell}\frac{\left(\nu_{n\ell}^{obs}-\nu_{n\ell}^{model}(\theta)\right)^2}{\sigma_{n\ell}^2}+\sum_k\lambda_k\mathcal{R}_k(\theta) \tag*{(90)}

where Rk\mathcal{R}_k encode constraints from spectroscopy, evolutionary tracks, and physical consistency.

Differentiable Formulation The key challenge is differentiating through the eigenvalue problem. We use the Hellmann-Feynman theorem:

∂ωnℓ2∂θi=⟨ξnℓ|∂L∂θi|ξnℓ⟩(91)\frac{\partial\omega_{n\ell}^2}{\partial\theta_i}=\left\langle\xi_{n\ell}\middle|\frac{\partial\mathcal{L}}{\partial\theta_i}\middle|\xi_{n\ell}\right\rangle \tag*{(91)}

where ∣ξnℓ⟩|\xi_{n\ell}\rangle is the normalized eigenfunction.

For the stellar structure, we solve the equations of hydrostatic equilibrium:

dmdr=4πr2ρ,dPdr=−Gmρr2,dLdr=4πr2ρε(92)\frac{dm}{dr}=4\pi r^2\rho,\qquad\frac{dP}{dr}=-\frac{Gm\rho}{r^2},\qquad\frac{dL}{dr}=4\pi r^2\rho\varepsilon \tag*{(92)}
dTdr=−3κρL16πacT3r2(radiative),dTdr=(1−1γ)TPdPdr(convective)(93)\frac{dT}{dr}=-\frac{3\kappa\rho L}{16\pi acT^3r^2}\quad\text{(radiative)},\qquad\frac{dT}{dr}=\left(1-\frac{1}{\gamma}\right)\frac{T}{P}\frac{dP}{dr}\quad\text{(convective)} \tag*{(93)}

These ODEs are differentiable, allowing gradients via the adjoint method:

∂L∂θ=∫0RλT(r)∂f∂θ(y(r),r,θ) dr(94)\frac{\partial\mathcal{L}}{\partial\theta}=\int_0^R\lambda^T(r)\frac{\partial f}{\partial\theta}(y(r),r,\theta)\,dr \tag*{(94)}

where λ(r)\lambda(r) is the adjoint variable satisfying a backward ODE.

Gradient Computation. We decompose the gradient calculation into three stages:

  1. Structure Derivatives: For parameter θi\theta_i, solve the linearized structure equations:

ddr(∂y∂θi)=∂f∂y∂y∂θi+∂f∂θi(95)\frac{d}{dr}\left(\frac{\partial\mathbf{y}}{\partial\theta_i}\right)=\frac{\partial\mathbf{f}}{\partial\mathbf{y}}\frac{\partial\mathbf{y}}{\partial\theta_i}+\frac{\partial\mathbf{f}}{\partial\theta_i} \tag*{(95)}

where y=(m,P,L,T)T\mathbf{y}=(m,P,L,T)^{\mathrm{T}} and f\mathbf{f} represents the RHS of the structure equations.

  1. Eigenvalue Derivatives: With perturbed structure, compute:

∂ωnℓ2∂θi=∫0Rc2r2[∂∂θi(ρω2−A)] dr(96)\frac{\partial\omega_{n\ell}^{2}}{\partial\theta_i}=\int_{0}^{R}\frac{c^{2}}{r^{2}}\left[\frac{\partial}{\partial\theta_i}\left(\rho\omega^{2}-A\right)\right]\,dr \tag*{(96)}

where AA contains the coefficient functions in the pulsation equations.

  1. Surface Correction: Account for near-surface errors:

νnℓcorrected=νnℓmodel+a0(νnℓνmax⁡)a1(97)\nu_{n\ell}^{\mathrm{corrected}}=\nu_{n\ell}^{\mathrm{model}}+a_{0}\left(\frac{\nu_{n\ell}}{\nu_{\max}}\right)^{a_{1}} \tag*{(97)}

with a0a_0, a1a_1 as additional parameters.

Algorithm: Differentiable Asteroseismic Inversion

Input: Observed frequencies {νnℓobs}\{\nu_{n\ell}^{\mathrm{obs}}\}, spectroscopic constraints (Teff,log⁡g,[Fe/H])(T_{\mathrm{eff}},\log g,[\mathrm{Fe}/\mathrm{H}])

Output: Stellar parameters θ∗\theta^{*} and internal structure

  1. Initialize with scaling relations: M≈(νmax⁡/νmax⁡,⊙)3(Δν/Δν⊙)−4M⊙M\approx(\nu_{\max}/\nu_{\max,\odot})^{3}(\Delta\nu/\Delta\nu_{\odot})^{-4}M_{\odot}

  2. Construct initial model using 1D stellar evolution code

  3. repeat a. Solve structure equations (m,P,L,T)(r)(m,P,L,T)(r) via shooting method b. Compute coefficient functions A(r)A(r), V(r)V(r), U(r)U(r) for pulsation c. Solve eigenvalue problem for frequencies {ωnℓ}\{\omega_{n\ell}\} d. Apply surface correction and mode identification e. Compute seismic χ2\chi^{2}:

χseis2=∑n,ℓ(νnℓobs−νnℓmodel)2σnℓ2\chi_{\mathrm{seis}}^{2}=\sum_{n,\ell}\frac{\left(\nu_{n\ell}^{\mathrm{obs}}-\nu_{n\ell}^{\mathrm{model}}\right)^{2}}{\sigma_{n\ell}^{2}}

f. Add spectroscopic constraints:

χspec2=(Teffobs−Teffmodel)2σT2+(log⁡gobs−log⁡gmodel)2σlog⁡g2\chi_{\mathrm{spec}}^{2}=\frac{\left(T_{\mathrm{eff}}^{\mathrm{obs}}-T_{\mathrm{eff}}^{\mathrm{model}}\right)^{2}}{\sigma_{T}^{2}}+\frac{\left(\log g^{\mathrm{obs}}-\log g^{\mathrm{model}}\right)^{2}}{\sigma_{\log g}^{2}}

g. Solve adjoint equations backward from surface h. Compute gradients via adjoint method i. Update: θ←θ−η∇θ(χseis2+λχspec2)\theta\leftarrow\theta-\eta\nabla_{\theta}(\chi_{\mathrm{seis}}^{2}+\lambda\chi_{\mathrm{spec}}^{2})

  1. until convergence or frequency residuals <0.1 μHz<0.1\,\mu\mathrm{Hz}

Numerical Considerations

  • Mode Identification: Use échelle diagram topology:

νnℓ≈Δν(n+ℓ2+ϵ)+δνnℓ(98)\nu_{n\ell}\approx\Delta\nu\left(n+\frac{\ell}{2}+\epsilon\right)+\delta\nu_{n\ell} \tag*{(98)}

Enforce ordering constraints: νn+1,ℓ>νn,ℓ\nu_{n+1,\ell}>\nu_{n,\ell}.

  • Adaptive Mesh: Concentrate grid points near discontinuities:

Δr∝∣d2ρdr2∣−1/3(99)\Delta r\propto\left|\frac{d^{2}\rho}{dr^{2}}\right|^{-1/3} \tag*{(99)}
  • Regularization: Penalize unphysical gradients:

Rsmooth=∫0R⋆∣d2ln⁡ρdr2∣2 dr(100)\mathcal{R}_{smooth} = \int_{0}^{R_\star} \left| \frac{d^2 \ln\rho}{d r^2} \right|^2 \, dr \tag*{(100)}

Exploration Strategies

  • Glitch Signatures: Parameterize sharp features:

δc2(r)=ABCZexp⁡(−(r−rBCZ)22w2)+AHeδ(r−rHe)(101)\delta c^2(r) = A_{BCZ} \exp\left(-\frac{(r-r_{BCZ})^2}{2w^2}\right) + A_{He}\delta(r-r_{He}) \tag*{(101)}

where rBCZr_{BCZ} and rHer_{He} locate the base of convection zone and helium ionization.

  • Rotation Inversion: Include rotational splitting:

νnℓm=νnℓ0+m∫0RKnℓ(r)Ω(r)dr(102)\nu_{n\ell m} = \nu_{n\ell0} + m \int_{0}^{R} K_{n\ell}(r)\Omega(r)dr \tag*{(102)}

Parameterize Ω(r)\Omega(r) with splines and jointly optimize.

  • Ensemble Analysis: For cluster members with common age/composition:

Lensemble=∑i=1Nstarsχi2+λage∑i<j(ti−tj)2+λZ∑i<j(Zi−Zj)2(103)\mathcal{L}_{ensemble} = \sum_{i=1}^{N_{stars}} \chi_i^2 + \lambda_{age}\sum_{i<j}(t_i-t_j)^2 + \lambda_Z\sum_{i<j}(Z_i-Z_j)^2 \tag*{(103)}

Radial Velocity Exoplanet Characterization

The radial velocity method [18] detects exoplanets through stellar wobble induced by gravitational interaction. Despite being mathematically smooth with well-defined derivatives, current analyses use Markov Chain Monte Carlo or genetic algorithms without gradient exploitation.

The stellar radial velocity due to NpN_p planets is:

vr(t)=∑p=1NpKp[cos⁡(fp(t)+ωp)+epcos⁡(ωp)]+vsys+vinst(t)(104)v_r(t) = \sum_{p=1}^{N_p} K_p[\cos(f_p(t)+\omega_p)+e_p\cos(\omega_p)] + v_{sys} + v_{inst}(t) \tag*{(104)}

where KpK_p is the semi-amplitude, fpf_p is the true anomaly, ωp\omega_p is the argument of periastron, and epe_p is eccentricity.

The true anomaly relates to time through Kepler’s equation:

M=E−esin⁡E,M=2πP(t−T0)(105)M = E - e\sin E,\qquad M = \frac{2\pi}{P}(t-T_0) \tag*{(105)}

where MM is mean anomaly and EE is eccentric anomaly.

Discovery Problem: Explore the parameter space θ={Kp,Pp,ep,ωp,T0,p}p=1Np\theta= \{K_p, P_p, e_p, \omega_p, T_{0,p}\}_{p=1}^{N_p} that minimizes:

L(θ)=∑i=1Nobs(vrobs(ti)−vrmodel(ti;θ))2σi2+R(θ)(106)\mathcal{L}(\theta) = \sum_{i=1}^{N_{obs}} \frac{(v_r^{obs}(t_i)-v_r^{model}(t_i;\theta))^2}{\sigma_i^2} + \mathcal{R}(\theta) \tag*{(106)}

discovering both the number of planets and their orbital configurations.

Differentiable Formulation The key challenge is differentiating through Kepler’s equation. We solve it iteratively:

En+1=M+esin⁡En(107)E_{n+1} = M + e\sin E_n \tag*{(107)}

converging to tolerance ∣En+1−En∣<ϵ\lvert E_{n+1}-E_n\rvert< \epsilon.

For gradients, we use implicit differentiation. At convergence, EE satisfies:

g(E,M,e)=E−esin⁡E−M=0(108)g(E,M,e) = E-e\sin E-M=0 \tag*{(108)}

By the implicit function theorem:

∂E∂M=11−ecos⁡E,∂E∂e=sin⁡E1−ecos⁡E(109)\frac{\partial E}{\partial M} = \frac{1}{1-e\cos E},\qquad\frac{\partial E}{\partial e} = \frac{\sin E}{1-e\cos E} \tag*{(109)}

The true anomaly is:

f=2arctan⁡(1+e1−etan⁡E2)(110)f = 2\arctan\left(\sqrt{\frac{1+e}{1-e}}\tan\frac{E}{2}\right) \tag*{(110)}

with derivatives following from the chain rule.

Gradient Computation For each orbital parameter, we trace gradients through the complete model:

  1. Period Gradient:

∂vr∂P=K∂∂P[cos⁡(f+ω)]=−Ksin⁡(f+ω)∂f∂P(111)\frac{\partial v_r}{\partial P}=K\frac{\partial}{\partial P}[\cos(f+\omega)]=-K\sin(f+\omega)\frac{\partial f}{\partial P} \tag*{(111)}

where:

∂f∂P=∂f∂E∂E∂M∂M∂P=∂f∂E11−ecos⁡E−2π(t−T0)P2(112)\frac{\partial f}{\partial P}=\frac{\partial f}{\partial E}\frac{\partial E}{\partial M}\frac{\partial M}{\partial P}=\frac{\partial f}{\partial E}\frac{1}{1-e\cos E}\frac{-2\pi(t-T_0)}{P^2} \tag*{(112)}
  1. Eccentricity Gradient: More complex due to appearance in multiple terms:

∂vr∂e=Kcos⁡ω+K∂∂e[cos⁡(f+ω)](113)\frac{\partial v_r}{\partial e}=K\cos\omega+K\frac{\partial}{\partial e}[\cos(f+\omega)] \tag*{(113)}
  1. Multi-Planet Coupling: Account for gravitational interactions:

Lstability=∑i<jexp⁡(−ΔaijRHill)(114)\mathcal{L}_{stability}=\sum_{i<j}\exp\left(-\frac{\Delta a_{ij}}{R_{Hill}}\right) \tag*{(114)}

Algorithm: Differentiable RV Exoplanet Discovery

Input: RV measurements {ti,vi,σi}\{t_i,v_i,\sigma_i\}, stellar mass M∗M_\ast Output: Number of planets Np∗N_p^\ast and orbital parameters θ∗\theta^\ast

  1. Initialize with periodogram peaks: identify NinitN_{init} candidate periods

  2. For Np=1N_p=1 to NmaxN_{max}: a. Initialize orbital parameters for NpN_p planets b. repeat i. Compute mean anomalies: Mp(ti)=2π(ti−T0,p)/PpM_p(t_i)=2\pi(t_i-T_{0,p})/P_p ii. Solve Kepler’s equation for Ep(ti)E_p(t_i) via Newton’s method iii. Calculate true anomalies: fp(ti)f_p(t_i) from Ep(ti)E_p(t_i) iv. Evaluate model: vrmodel(ti)=∑pKp[cos⁡(fp+ωp)+epcos⁡ωp]v_r^{model}(t_i)=\sum_p K_p[\cos(f_p+\omega_p)+e_p\cos\omega_p] v. Compute residuals and χ2\chi^2 vi. Add regularization: - Eccentricity prior: Re=∑pβep2R_e=\sum_p\beta e_p^2 - Stability constraint: RstabR_{stab} vii. Compute gradients via automatic differentiation viii. Update: θ←θ−η∇θL\theta\leftarrow\theta-\eta\nabla_\theta\mathcal{L} c. until convergence d. Compute BIC: BIC=χ2+kln⁡NobsBIC=\chi^2+k\ln N_{obs} where k=5Npk=5N_p

  3. Select Np∗N_p^\ast minimizing BIC

  4. Refine with activity modeling: vactivity(t)=∑kAksin⁡(2πt/Prot+ϕk)v_{activity}(t)=\sum_k A_k\sin(2\pi t/P_{rot}+\phi_k)

Numerical Considerations

  • Kepler Solver Stability: Use Markley’s method for e>0.95e>0.95:

E=M+esin⁡M1+ecos⁡M1−e2+ecos⁡M(115)E=M+e\sin M\frac{1+e\cos M}{1-e^2+e\cos M} \tag*{(115)}
  • Period Aliasing: Search periods on logarithmic grid:

Pk=Pminexp⁡(kln⁡(Pmax/Pmin)Ngrid)(116)P_k=P_{min}\exp\left(k\frac{\ln(P_{max}/P_{min})}{N_{grid}}\right) \tag*{(116)}
  • Stellar Activity: Model quasi-periodic variations:

vGP(t)∼GP(0,k(t,t′)) where k=Aexp⁡(−∣t−t′∣22l2−Γsin⁡2π∣t−t′∣Prot)(117)v_{GP}(t)\sim\mathcal{GP}(0,k(t,t'))\ \text{where}\ k=A\exp\left(-\frac{\lvert t-t'\rvert^2}{2l^2}-\Gamma\sin^2\frac{\pi\lvert t-t'\rvert}{P_{rot}}\right) \tag*{(117)}

Exploration Strategies

  • Hierarchical Discovery: Start with strongest signal, progressively add planets:

θNp+1={θNp∗,θnew}(118)\theta_{N_{p+1}} = \{\theta_{N_p}^{*}, \theta_{\mathrm{new}}\} \tag*{(118)}
  • Transit-RV Joint Fit: For transiting systems:

Ljoint=χRV2+χtransit2+λ∣ρ∗−ρmodel∣2(119)\mathcal{L}_{\mathrm{joint}} = \chi_{\mathrm{RV}}^{2} + \chi_{\mathrm{transit}}^{2} + \lambda\lvert\rho_{*} - \rho_{\mathrm{model}} \rvert^{2} \tag*{(119)}

where stellar density ρ∗\rho_{*} links both datasets.

  • Population-Level Inference: Hierarchical Bayesian approach:

θi∼p(θ∣ϕ),ϕ∼p(ϕ)(120)\theta_i \sim p(\theta\mid\phi), \quad\phi\sim p(\phi) \tag*{(120)}

where ϕ\phi are hyperparameters describing the population.

Asteroid Thermophysical Characterization

The Yarkovsky effect [19] causes orbital drift in asteroids due to anisotropic thermal emission. This subtle force depends on thermal inertia, rotation state, and surface properties—parameters currently constrained through computationally expensive Monte Carlo methods.

The surface temperature T(θ,ϕ,t)T(\theta,\phi,t) satisfies the heat–diffusion equation:

∂T∂t=κ∇2T,\frac{\partial T}{\partial t} = \kappa\nabla^{2}T,
ρCp∂T∂t=1r2∂∂r(r2κ∂T∂r)(121)\rho C_{p}\frac{\partial T}{\partial t} = \frac{1}{r^{2}}\frac{\partial}{\partial r}\left(r^{2}\kappa\frac{\partial T}{\partial r}\right) \tag*{(121)}

with boundary condition:

κ∂T∂r∣r=R=(1−A)F⊙cos⁡ξ−ϵσT4(122)\left.\kappa\frac{\partial T}{\partial r}\right|_{r=R} = (1-A)F_{\odot}\cos\xi- \epsilon\sigma T^{4} \tag*{(122)}

where ξ\xi is the solar zenith angle and F⊙F_{\odot} is the solar flux.

The recoil acceleration from thermal emission is:

aYark=−23mc∫SϵσT4n^ dS(123)\mathbf{a}_{Yark} = -\frac{2}{3mc}\int_{S}\epsilon\sigma T^{4}\hat{\mathbf{n}}\,dS \tag*{(123)}

Over many orbits, this produces secular drift:

dadt=2naaYark⋅t^(124)\frac{da}{dt} = \frac{2}{na}\mathbf{a}_{Yark}\cdot\hat{\mathbf{t}} \tag*{(124)}

where t^\hat{\mathbf{t}} is the transverse direction.

Discovery Problem: Given observed orbital drift (da/dt)obs(da/dt)_{\mathrm{obs}} from astrometry, discover thermal and physical parameters θ={Γ,ρ,Cp,A,ϵ,Prot,γ}\theta= \{\Gamma,\rho,C_{p},A,\epsilon,P_{\mathrm{rot}},\gamma\} that minimize:

L(θ)=((da/dt)obs−(da/dt)model(θ)σda/dt)2+∑kλkRk(θ)(125)\mathcal{L}(\theta) = \left(\frac{(da/dt)_{\mathrm{obs}} - (da/dt)_{\mathrm{model}}(\theta)}{\sigma_{da/dt}}\right)^{2} + \sum_{k}\lambda_{k}R_{k}(\theta) \tag*{(125)}

Differentiable Formulation We discretize the heat equation using finite differences in depth and spherical harmonics on the surface:

T(r,θ,ϕ,t)=∑ℓ=0Lmax⁡∑m=−ℓℓTℓm(r,t)Yℓm(θ,ϕ)(126)T(r,\theta,\phi,t) = \sum_{\ell=0}^{L_{\max}}\sum_{m=-\ell}^{\ell}T_{\ell m}(r,t)Y_{\ell m}(\theta,\phi) \tag*{(126)}

The time evolution becomes:

∂Tℓm∂t=κρCpr2∂∂r(r2∂Tℓm∂r)(127)\frac{\partial T_{\ell m}}{\partial t} = \frac{\kappa}{\rho C_{p}r^{2}}\frac{\partial}{\partial r}\left(r^{2}\frac{\partial T_{\ell m}}{\partial r}\right) \tag*{(127)}

For the surface heating, we expand:

cos⁡ξ=∑ℓ,mSℓm(t)Yℓm(θ,ϕ)(128)\cos\xi= \sum_{\ell,m}S_{\ell m}(t)Y_{\ell m}(\theta,\phi) \tag*{(128)}

where SℓmS_{\ell m} encode the solar direction in the body-fixed frame.

Gradient Computation The gradient flow requires:

  1. Temperature Sensitivity: Solve the linearized heat equation:

∂∂t(∂T∂θi)=1ρCp∇⋅(κ∇∂T∂θi)+∂∂θi(1ρCp∇⋅(κ∇T))(129)\frac{\partial}{\partial t}\left(\frac{\partial T}{\partial\theta_i}\right)=\frac{1}{\rho C_p}\nabla\cdot\left(\kappa\nabla\frac{\partial T}{\partial\theta_i}\right)+\frac{\partial}{\partial\theta_i}\left(\frac{1}{\rho C_p}\nabla\cdot(\kappa\nabla T)\right) \tag*{(129)}
  1. Force Sensitivity: Integrate over the surface:

∂aYark∂θi=−8ϵσ3mc∫ST3∂T∂θin^ dS(130)\frac{\partial\mathbf{a}_{Yark}}{\partial\theta_i}=-\frac{8\epsilon\sigma}{3mc}\int_S T^3\frac{\partial T}{\partial\theta_i}\hat{\mathbf{n}}\,dS \tag*{(130)}
  1. Orbital Evolution: Propagate through Gauss’s equations:

∂∂θi(dadt)=2na∂aYark∂θi⋅t^+2naaYark⋅∂t^∂θi∂a∂θi(131)\frac{\partial}{\partial\theta_i}\left(\frac{da}{dt}\right)=\frac{2}{na}\frac{\partial\mathbf{a}_{Yark}}{\partial\theta_i}\cdot\hat{\mathbf{t}}+\frac{2}{na}\mathbf{a}_{Yark}\cdot\frac{\partial\hat{\mathbf{t}}}{\partial\theta_i}\frac{\partial a}{\partial\theta_i} \tag*{(131)}

Algorithm: Differentiable Yarkovsky Effect Inversion

Input: Orbital drift (da/dt)obs(da/dt)_{obs}, shape model, orbital elements Output: Thermal parameters {Γ,ρCp}\{\Gamma,\rho C_p\} and spin state

  1. Initialize thermal inertia Γ=100 J m−2 K−1 s−1/2\Gamma= 100\ \mathrm{J\ m^{-2}\ K^{-1}\ s^{-1/2}}

  2. Discretize: radial grid {ri}\{r_i\}, surface harmonics YℓmY_{\ell m}

  3. repeat a. Solve heat equation over one rotation:

    • Compute insolation pattern Sℓm(t)S_{\ell m}(t)

    • Time-step temperature field Tℓm(ri,t)T_{\ell m}(r_i,t)

    • Apply surface boundary conditions b. Calculate thermal force:

    • Surface temperatures: Ts(θ,ϕ)=∑Tℓm(R)YℓmT_s(\theta,\phi)=\sum T_{\ell m}(R)Y_{\ell m}

    • Emission pattern: Femit=ϵσTs4F_{\mathrm{emit}}=\epsilon\sigma T_s^4

    • Recoil force: aYark=−23mc∫Femitn^ dS\mathbf{a}_{Yark}=-\frac{2}{3mc}\int F_{\mathrm{emit}}\hat{\mathbf{n}}\,dS c. Propagate orbital evolution:

    • Average over rotation: ⟨aYark⟩\langle\mathbf{a}_{Yark}\rangle

    • Compute secular drift: (da/dt)model(da/dt)_{\mathrm{model}} d. Evaluate loss with physical constraints:

L=((da/dt)obs−(da/dt)modelσ)2+λ1(Γ−Γprior)2\mathcal{L}=\left(\frac{(da/dt)_{obs}-(da/dt)_{\mathrm{model}}}{\sigma}\right)^2+\lambda_1(\Gamma-\Gamma_{prior})^2

e. Compute gradients via adjoint method through heat equation f. Update: θ←θ−η∇θL\theta\leftarrow\theta-\eta\nabla_\theta\mathcal{L}

  1. until convergence

Numerical Considerations

  • Skin Depth Resolution: Radial grid spacing near surface:

Δrmin<12κProtπρCp(132)\Delta r_{min}<\frac{1}{2}\sqrt{\frac{\kappa P_{rot}}{\pi\rho C_p}} \tag*{(132)}
  • Spherical Harmonic Truncation: Choose LmaxL_{max} based on shape complexity:

Lmax≈2πRλmin(133)L_{max}\approx\frac{2\pi R}{\lambda_{min}} \tag*{(133)}

where λmin\lambda_{min} is the smallest surface feature.

  • Rotation State: Parameterize using quaternions to avoid singularities:

q(t)=exp⁡(t2[ωxi+ωyj+ωzk])q0(134)\mathbf{q}(t) = \exp\left(\frac{t}{2}\left[\omega_x\mathbf{i}+\omega_y\mathbf{j}+\omega_z\mathbf{k}\right]\right)\mathbf{q}_0 \tag*{(134)}

Exploration Strategies

  • Multi-Wavelength Constraints: Include thermal IR observations:

Ltotal=Lorbit+∑λ(Fλobs−Fλmodel(θ))2σλ2(135)\mathcal{L}_{total} = \mathcal{L}_{orbit} + \sum_{\lambda}\frac{\left(F_{\lambda}^{obs}-F_{\lambda}^{model}(\theta)\right)^2}{\sigma_{\lambda}^2} \tag*{(135)}
  • Binary Asteroids: Model mutual heating:

κ∂T1∂r∣surface=(1−A1)F⊙cos⁡ξ1+Fmutual−ϵ1σT14(136)\left.\kappa\frac{\partial T_1}{\partial r}\right|_{surface} = (1-A_1)F_{\odot}\cos\xi_1+F_{mutual}-\epsilon_1\sigma T_1^4 \tag*{(136)}

where FmutualF_{mutual} is thermal radiation from the companion.

  • YORP Effect: Simultaneous fit of orbital and rotational evolution:

dωdt=1I∫Sr×(n^×Femit) dS(137)\frac{d\boldsymbol{\omega}}{dt}=\frac{1}{I}\int_S\mathbf{r}\times(\hat{\mathbf{n}}\times\mathbf{F}_{emit})\,dS \tag*{(137)}
  • Population Studies: Hierarchical model for asteroid families:

Γi∼log⁡N(μΓ,σΓ2),optimize {μΓ,σΓ}(138)\Gamma_i\sim\log\mathcal{N}(\mu_{\Gamma},\sigma_{\Gamma}^2),\quad\text{optimize }\{\mu_{\Gamma},\sigma_{\Gamma}\} \tag*{(138)}

The differentiable formulation transforms asteroid thermophysical characterization from a parameter estimation problem to a discovery framework, revealing connections between thermal properties, rotation states, and orbital evolution that traditional methods miss due to computational limitations.

Discussion

Implementation Considerations

The mathematical foundations presented across all nine applications demonstrate that each domain involves inherently differentiable operations amenable to gradient-based exploration. Modern AD frameworks elegantly handle the diverse technical challenges encountered:

  • Implicit Differentiation: Fixed-point problems (Kepler’s equation in RV fitting, bounce solutions in false vacuum decay)

  • ODE/PDE Adjoints: Continuous systems (stellar structure equations, heat diffusion for Yarkovsky effect, string worldsheet evolution)

  • Complex Gradients: Wirtinger calculus for interferometric calibration and visibility modeling

  • Eigenvalue Sensitivity: Hellmann-Feynman theorem for asteroseismic frequencies

  • Memory-Efficient Backpropagation: High-resolution black hole imaging and 3D spacetime metrics

  • Constraint Manifolds: Projected gradients for physical constraints (positivity, gauge fixing, boundary conditions)

The consistent appearance of these patterns across disparate astrophysical scales—from quantum bubble nucleation to galactic-scale cosmic strings—suggests that differentiable discovery represents a unifying computational paradigm for theoretical astrophysics.

The Discovery Paradigm: Beyond Optimization

The transformation from optimization to discovery merits emphasis. Traditional approaches in each domain typically involve:

  1. Human-designed parameterizations based on physical intuition

  2. Local optimization around expected solutions

  1. Limited exploration due to computational constraints

GRASP inverts this workflow:

  1. Neural or basis function parameterizations that span all possible configurations

  2. Global exploration through multi-start gradient descent

  3. Comprehensive mapping enabled by 100-1000× computational efficiency gains

This paradigm shift has profound implications. In the Alcubierre metric optimization, neural parameterization of the shape function could discover spacetime geometries that violate human geometric intuition while better satisfying energy conditions. For asteroseismology, simultaneous optimization of stellar parameters and mode identification could reveal previously hidden relationships between internal structure and oscillation patterns.

GRASP Framework Vision

We envision GRASP not as a monolithic library implementing all nine domains, but as a conceptual framework and set of design principles for making astrophysical computations differentiable. The framework would provide:

  • Common patterns: Template implementations for typical astrophysical operations:

    • Complex gradient computation via Wirtinger calculus

    • ODE/PDE integration with adjoint sensitivity

    • Constraint handling through augmented Lagrangians

    • Multi-objective Pareto frontier exploration

  • Example implementations: Full reference implementations for 2-3 domains (e.g., interferometric calibration and asteroseismic inversion)

  • Best practices: Guidelines for:

    • Numerical stability in extreme parameter regimes

    • Efficient parameterization of physical fields

    • Adaptive discretization strategies

    • Regularization design for ill-posed problems

  • Interface standards: Common APIs enabling interoperability:

    • Forward model specification

    • Constraint definition

    • Loss function composition

    • Solution space visualization

Individual research groups with domain expertise would implement specific applications. For instance:

import grasp
import torch

# Example: A research group implements asteroseismology
# using GRASP design patterns
class StellarModel(grasp.DifferentiableModel):
    def __init__(self):
        self.mass = grasp.Parameter(1.0, bounds=(0.8, 1.2))
        self.metallicity = grasp.Parameter(0.0, bounds=(-0.5, 0.5))
        self.helium = grasp.Parameter(0.28, bounds=(0.24, 0.32))

    def forward(self):
        # Solve stellar structure equations
        structure = grasp.ode.solve_bvp(
            self.structure_equations,
            self.boundary_conditions
        )
# Compute oscillation frequencies
frequencies = grasp.eigenvalue.solve(
    self.pulsation_operator(structure)
)
return frequencies

# Discovery through multi-start exploration
explorer = grasp.Explorer(
    model=StellarModel(),
    loss_fn=grasp.losses.ChiSquared(observed_frequencies),
    n_starts=100,
    device=’cuda’
)

# Discover all stellar configurations matching observations
solutions = explorer.discover()

# Visualize the solution landscape
grasp.visualize.parameter_space(solutions)

This distributed approach leverages community expertise while maintaining consistent design principles across domains.

Computational Advantages and Scaling

Beyond mathematical elegance, gradient-based discovery offers transformative practical advantages:

  1. Efficiency Gains:

    • One gradient evaluation approx. 2-3 forward evaluations (via reverse-mode AD)

    • Population methods: 100-1000 forward evaluations per meaningful update

    • Net speedup: 50-500× per optimization iteration

  2. Accuracy: Machine-precision gradients eliminate:

    • Finite difference truncation errors

    • Numerical noise in derivative estimates

    • Step-size tuning complications

  3. Scalability:

    • Reverse-mode AD cost independent of parameter dimension

    • Million-dimensional optimizations become feasible

    • Natural problem decomposition for distributed computing

  4. Hardware Acceleration:

    • Native GPU support in PyTorch/JAX

    • Automatic kernel fusion and optimization

    • Seamless scaling from laptops to supercomputers

Limitations and Future Directions

While the mathematical frameworks are complete, several practical considerations arise:

  • Discrete Parameters: Some applications involve integers (e.g., number of planets) requiring relaxation strategies

  • Extreme Parameter Regimes: Numerical stability near singularities (e.g., r→0r \to0 in GR applications)

  • Computational Singularities: Cusps in cosmic strings, coordinate singularities in black hole metrics

  • Scale Disparities: Stellar evolution timescales vs. oscillation periods require careful multi-scale methods

The immediate future work is clear: implement these nine applications. Each domain presents unique opportunities for discovery:

  • Alcubierre Optimization: First systematic discovery of minimal exotic matter warp configurations

  • Cosmic String Networks: Constrain string parameters from gravitational wave observations

  • Black Hole Imaging: Next-generation EHT reconstructions with optimized regularization

  • Asteroseismology: Systematic inversion for PLATO and TESS targets

  • RV Exoplanets: Unified framework handling systems with 10 or more planets

  • Yarkovsky Effect: Population-level thermal property constraints for asteroid families

We hope this paper inspires domain experts to implement these methods, potentially discovering new physics hiding in unexplored parameter spaces.

Conclusion

We have identified nine astrophysical domains where automatic differentiation remains unexplored for systematic discovery of solution spaces. Building on GRAF’s success in discovering optimal radar waveforms through gradient-based exploration, GRASP extends this paradigm to astrophysics, transforming optimization problems into comprehensive exploration of physically valid solution landscapes.

GRASP’s power lies in its ability to discover rather than merely optimize. Where traditional methods sample parameter spaces sparsely due to computational constraints, GRASP’s automatic differentiation foundation enables systematic mapping of complete solution landscapes, naturally leveraging modern GPU capabilities.

The mathematical foundations presented here show that each domain admits differentiable formulations for systematic exploration. Modern AD frameworks provide the infrastructure to discover not just local optima, but the full spectrum of physically permitted solutions. This establishes a new paradigm: instead of relying on initial guesses or incremental refinements, GRASP systematically discovers what the universe fundamentally permits. The most physically meaningful solutions emerge organically from the complete parameter space —often in regions no human or traditional algorithm would think to explore.

While GRASP currently provides a theoretical scaffolding, its primary contribution is establishing the viability and urgency of gradient-based discovery as a new computational paradigm in astrophysics.

Data and Software Availability

GRASP represents a framework and methodology rather than a single codebase. Example implementations demonstrating key concepts will be made available. We encourage domain experts to develop and share implementations for their specific applications, following the design principles outlined in this work.

Acknowledgments

We thank the developers of PyTorch, JAX, and TensorFlow for creating and maintaining the automatic differentiation frameworks that enable this research direction.

References

  1. [1]Paszke, A., et al., “PyTorch: An Imperative Style, High-Performance Deep Learning Library”, Advances in Neural Information Processing Systems, vol. 32, pp. 8024-8035, 2019.arxiv.org/abs/1912.01703
  2. [2]Bradbury, J., et al., “JAX: composable transformations of Python+NumPy programs”, 2018. [Online]. Available: http://github.com/google/jax
  3. [3]Abadi, M., et al., “TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems”, 2015. Software available from tensorflow.orgarxiv.org/abs/1603.04467
  4. [4]M. Bara, “Differentiable Radar Ambiguity Functions: Mathematical Formulation and Computational Implementation”, arXiv preprint arXiv:2506.22935, 2025.
  5. [5]W. Doyle, “Searching For Black Holes Using Auto Differentiation”, Emerging Minds Journal for Student Research, vol. 1, pp. 17-38, 2023.DOI
  6. [6]T. D. P. Edwards, K. W. K. Wong, K. Jani, T. Wouters, S. M. Schäfer, and J. Gair, “ripple: Differentiable and Hardware-Accelerated Waveforms for Gravitational Wave Data Analysis”, arXiv preprint arXiv:2302.05329, 2023.arxiv.org/abs/2302.05329
  7. [7]M. Chianese, A. Coogan, P. Hofma, S. Otten, and C. Weniger, “Differentiable Strong Lensing: Uniting Gravity and Neural Nets through Differentiable Probabilistic Programming”, Monthly Notices of the Royal Astronomical Society, vol. 496, no. 1, pp. 381-393, 2020.DOI
  8. [8]J.-E. Campagne, et al., “JAX-COSMO: An End-to-End Differentiable and GPU Accelerated Cosmology Library”, arXiv preprint arXiv:2302.05163, 2023.DOI
  9. [9]H. Kawahara, Y. Kawashima, K. Masuda, I. J. M. Crossfield, E. Pannier, and D. C. van den Bekerom, “Autodifferentiable Spectrum Model for High-dispersion Characterization of Exoplanets and Brown Dwarfs”, The Astrophysical Journal Supplement Series, vol. 258, no. 2, p. 31, 2022.DOI
  10. [10]S. Hattori, L. Garcia, C. Murray, J. Dong, S. Dholakia, D. Degen, and D. Foreman-Mackey, “exoplanet-dev/jaxplanet: Astronomical time series analysis with JAX (v0.0.2)”, Zenodo, 2024.
  11. [11]M. Alcubierre, “The warp drive: hyper-fast travel within general relativity”, Classical and Quantum Gravity, vol. 11, no. 5, pp. L73-L77, 1994.DOI
  12. [12]A. Vilenkin and E. P. S. Shellard, Cosmic Strings and Other Topological Defects. Cambridge University Press, 2000.
  13. [14]S. Coleman, “Fate of the false vacuum. I. Semiclassical theory”, Physical Review D, vol. 15, no. 10, pp. 2929-2936, 1977.DOI
  14. [15]A. R. Thompson, J. M. Moran, and G. W. Swenson Jr., Interferometry and Synthesis in Radio Astronomy, 3rd ed. Springer, 2017.
  15. [16]The Event Horizon Telescope Collaboration, “First M87 Event Horizon Telescope Results. I. The Shadow of the Supermassive Black Hole”, The Astrophysical Journal Letters, vol. 875, p. L1, 2019.DOI
  16. [17]C. Aerts, J. Christensen-Dalsgaard, and D. W. Kurtz, Asteroseismology. Springer, 2010.DOI
  17. [18]M. Mayor and D. Queloz, “A Jupiter-mass companion to a solar-type star”, Nature, vol. 378, pp. 355-359, 1995.DOI
  18. [19]W. F. Bottke Jr., D. Vokrouhlický, D. P. Rubincam, and D. Nesvorný, “The Yarkovsky and YORP Effects: Implications for Asteroid Dynamics”, Annual Review of Earth and Planetary Sciences, vol. 34, pp. 157-191, 2006.DOI
  19. [20]A. Vilenkin and E. P. S. Shellard, Cosmic Strings and Other Topological Defects. Cambridge University Press, 2000.
  20. [21]A. Lopez-Eiguren, J. Urrestilla, M. Hindmarsh, J. C. Bueno Sanchez, D. Daverio, and M. Kunz, “Cosmic Microwave Background constraints for global strings and global monopoles”, Journal of Cosmology and Astroparticle Physics, vol. 2017, no. 07, p. 026, 2017.
  21. [22]G. E. Freedman, Low-Frequency Gravitational Wave Searches and Data Analysis with Hamiltonian Sampling, Ph.D. Dissertation, University of Wisconsin-Milwaukee, 2025.
  22. [23]Utkarsh Giri, Bridget C. Andersen, Pragya Chawla, Alice P. Curtin, Emmanuel Fonseca, Victoria M. Kaspi, Hsiu-Hsien Lin, Kiyoshi W. Masui, Ketan R. Sand, Paul Scholz, Thomas C. Abbott, Fengqiu Adam Dong, B. M. Gaensler, Calvin Leung, Daniele Michilli, Mohit Bhardwaj, Moritz Münchmeyer, Ayush Pandhi, Aaron B. Pearlman, Ziggy Pleunis, Masoud Rafiei-Ravandi, Alex Reda, Kaitlyn Shin, Kendrick Smith, Ingrid H. Stairs, David C. Stenning, Shriharsh P. Tendulkar, “Comprehensive Bayesian analysis of FRB-like bursts from SGR 1935+2154 observed by CHIME/FRB”, arXiv preprint arXiv:2310.16932, 2023.arxiv.org/abs/2310.16932
  23. [24]Jennifer R. Andersson, Oleg Kochukhov, Zheng Zhao, and Jens Sjölund, “Bayesian Uncertainty Quantification in High-dimensional Stellar Magnetic Field Models,” ICML 2023 Workshop on Machine Learning for Astrophysics, PMLR 202, 2023.
  24. [25]G. Singh, J. Kaur, G. Singh, H. Singh, J. Gill, R. Singh, and M. Singh, “Artificial Intelligence & Quantum Computing for Causal Loop and Chronology Violation Detection: A Deep Learning Approach to Predicting Traversable Wormhole Stability and Temporal Paradoxes,” International Journal of Innovative Scientific Research, vol. 3, no. 2, pp. 1–13, 2025.

Paper details

vlnc:W_bcnyc3ps
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents