betaFeedback

WorksMathematics

The Reconstruction of the Space-Dependent Thermal Conductivity from Sparse Temperature Measurements

Abstract

We present a novel method for reconstructing the thermal conductivity coefficient in 1D and 2D heat equations using moving sensors that dynamically traverse the domain to record sparse and noisy temperature measurements. We significantly reduce the computational cost associated with forward PDE evaluations by employing automatic differentiation, enabling a more efficient and scalable reconstruction process. This allows the inverse problem to be solved with fewer sensors and observations. Specifically, we demonstrate the successful reconstruction of thermal conductivity on the 1D circle and 2D torus, using one and four moving sensors, respectively, with their positions recorded over time. Our method incorporates sampling algorithms to compute confidence intervals for the reconstructed conductivity, improving robustness against measurement noise. Extensive numerical simulations of heat dynamics validate the efficacy of our approach, confirming both the accuracy and stability of the reconstructed thermal conductivity. Additionally, the method is thoroughly tested using large datasets from machine learning, allowing us to evaluate its performance across various scenarios and ensure its reliability. This approach provides a cost-effective and flexible solution for conductivity reconstruction from sparse measurements, making it a robust tool for solving inverse problems in complex domains.

Introduction

This paper addresses the numerical solution of the inverse heat equation, specifically focusing on reconstructing thermal conductivity from temperature sensor measurements. Thermal conductivity, a critical property governing heat diffusion, is essential in various industrial applications, including cooling of microchips [9], temperature regulation in chemical reactors, and the optimization of heat-dissipating devices [4]. These applications require precise temperature control to ensure the efficiency, safety, and longevity of the components involved.

While thermal conductivity values for homogeneous materials like metals or liquids are well-documented [10], measuring conductivity in inhomogeneous materials with spatially varying properties poses a significant challenge. Traditional methods often require physically breaking down the material into smaller, homogeneous pieces, which can compromise the material’s integrity and yield inaccurate measurements.

Recent advancements in sensor technology have enabled the development of non-invasive methods to overcome these limitations. Stuart [2010] provides a method within a Bayesian framework and applies it to reconstruct the thermal conductivity of a one-dimensional rod. However, this approach requires many sensors on the rod to achieve accurate reconstruction. Beskos et al. [2017] extends the Bayesian framework to two-dimensional non-convex domains, but the requirement for numerous sensors remains. Zhang et al. [2013] proposed the homotopy regularization algorithm with sigmoid-type homotopy parameter to simultaneously invert the space-dependent diffusion coefficient and source magnitude, but it works only for one-dimensional spatial domain. More recently, Boumenir [2023] proposed a promising alternative method using multiple temperature measurements with only a single sensor to accurately reconstruct the thermal conductivity of a one-dimensional rod. However, it relies on multiple physical experiments with different initial conditions, which can be costly and time-consuming.

To address these challenges, we propose a novel method utilizing moving sensors that dynamically traverse the domain while recording both temperature and position data. In one-dimension domains, this simply means the sensor is sliding on the rod. This approach fully utilizes the sensor’s ability to record temperature over time, forming a time series of temperature and position measurements. This enables the effective reconstruction of the thermal conductivity from a single physical experiment, completely eliminating the need for multiple experiments and significantly reduces the number of sensors. By leveraging machine learning techniques and automatic differentiation, we optimize the reconstruction process and minimize the computational costs associated with forward evaluations of partial differential equations (PDEs). These advancements make our method efficient, scalable, and practical for real-world applications. Furthermore, we extend this method to two-dimensional domains, such as the two-dimensional torus, where fewer moving sensors are required to achieve accurate reconstructions, thus lowering overall costs. A practical application could involve reconstructing the thermal conductivity of a city using temperature sensors mounted on vehicles, with GPS tracking providing accurate positional data.

We rigorously validate our method through numerical simulations and large-scale testing on machine learning datasets, ensuring its robustness across various scenarios. Our theoretical contributions include a thorough analysis of numerical stability, leveraging the spectral properties of the underlying problem to ensure robustness against errors in the conductivity reconstruction. Additionally, this work builds on and extends the research previously published by Boumenir [2023] and Rashedi [2021] further advancing techniques in thermal conductivity reconstruction for inhomogeneous materials. Given the significance of this method, we believe it offers a cost-effective and practical solution for a wide range of applications, demonstrating its potential for broader use in real-world scenarios.

Problem background

We consider thermal conductivity recovery from sparse temperature measurements. For conciseness, all mentions of conductivity in the following context refer to thermal conductivity. The underlying dynamics of temperature follows

ut−∇⋅[a(x)∇u]=f(t,x),(1)u_t-\nabla\cdot\left[a(\mathbf{x})\nabla u\right]=f(t,\mathbf{x}), \tag*{(1)}

subject to periodic boundary conditions. Here, a:D→R+a:\mathcal{D}\to\mathbb{R}^{+} denotes the conductivity, which is constant in both time and temperature [Zheng et al., 2011], and u:R+×D∋(t,x)→Ru:\mathbb{R}^{+}\times\mathcal{D}\ni(t,\mathbf{x})\to\mathbb{R} denotes the temperature. The initial condition u(t=0,x)u(t=0,\mathbf{x}) on the domain is given. For simplicity, throughout this work we consider the spatial domain D=T=R/Z\mathcal{D}=\mathbb{T}=\mathbb{R}/\mathbb{Z} in the 1D case or D=T2=R2/Z2\mathcal{D}=\mathbb{T}^{2}=\mathbb{R}^{2}/\mathbb{Z}^{2} in the 2D case, with the domain represented as [0,1)[0,1) or [0,1)2[0,1)^{2}, respectively. We restrict our focus to periodic boundary conditions, leaving other boundary conditions such as Dirichlet or Neumann for future consideration. In the 1D case, the gradient operator is ∇=∂x\nabla=\partial_x, while in the 2D case,

∇=[∂x∂y]T,\nabla=\begin{bmatrix}\partial_x & \partial_y\end{bmatrix}^{\mathsf{T}},

denoting the spatial derivatives.

Recoverability

We first rule out the scenarios where it is impossible to recover the conductivity from temperature measurements due to the lack of uniqueness. We propose the following criteria where even full observation of temperature fails to recover any information about a(x)a(\mathbf{x}).

Proposition 2.1 (Non-recoverability of conductivity). Consider the temperature dynamics governed by (1) with the initial condition u0(x):=u(t=0,x)u_0(\mathbf{x}):=u(t=0,\mathbf{x}) provided. If the following conditions hold:

  • ∇u(t=0,x)=0∀x∈D\nabla u(t=0,\mathbf{x})=0\quad\forall\mathbf{x}\in\mathcal{D},

  • ∇f(t,x)=0∀(t,x)∈R+×D\nabla f(t,\mathbf{x})=0\quad\forall(t,\mathbf{x})\in\mathbb{R}^{+}\times\mathcal{D},

then the conductivity a(x)a(\mathbf{x}) cannot be reconstructed from {u(t,x)}R+×D\{u(t,\mathbf{x})\}_{\mathbb{R}^{+}\times\mathcal{D}}, which is the full temperature observation.

Proof. Given ∇u(t=0,x)=0\nabla u(t=0,\mathbf{x})=0 and ∇f(t,x)=0\nabla f(t,\mathbf{x})=0, (1) simplifies to:

ut=f(t),where u(t,x)=u(t) and f(t,x)=f(t),∀(t,x)∈R+×D,u_t=f(t),\qquad\text{where }u(t,\mathbf{x})=u(t)\text{ and }f(t,\mathbf{x})=f(t),\quad\forall(t,\mathbf{x})\in\mathbb{R}^{+}\times\mathcal{D},

which is a scalar ordinary differential equation (ODE) in time, independent of x\mathbf{x}. The solution to this ODE is:

u(t)=u(0)+∫0tf(s) ds.u(t)=u(0)+\int_{0}^{t}f(s)\,\mathrm{d}s.

Since a(x)a(\mathbf{x}) does not appear in this solution, it cannot be reconstructed from the temperature observations.

Remark. The condition ∇u(t=0,x)=0\nabla u(t=0,\mathbf{x})=0 indicates that the initial temperature distribution is spatially constant, preventing any heat flow between locations. Additionally, ∇f(t,x)=0\nabla f(t,\mathbf{x})=0 ensures that the external heat source is uniform across the domain, so the temperature distribution remains spatially constant throughout the time. These conditions lead to a degenerate scenario where (1) reduces to an ODE, and the spatial variation of conductivity a(x)a(\mathbf{x}) becomes irrelevant, resulting in the loss of uniqueness in the inverse problem.

Lemma 2.2 (Lipschitz dependence of solution on conductivity). Let domain be D=T\mathcal{D} = \mathbb{T} or D=T2\mathcal{D} = \mathbb{T}^2. Let a,b∈C∞(D;R+)a,b \in C^\infty(\mathcal{D};\mathbb{R}^+) be two bounded non-degenerate conductivities:

m:=min⁡{inf⁡Da,inf⁡Db}>0,M:=sup⁡D{a+b}<∞.m := \min\left\{\inf_{\mathcal{D}} a,\inf_{\mathcal{D}} b\right\} > 0,\qquad M := \sup_{\mathcal{D}}\{a+b\} < \infty.

Let u(t,x;a),u(t,x;b)u(t,x;a),u(t,x;b) denote the solution to (1) with the same inhomogeneous term ff and the same initial condition

u0(⋅):=u(t=0,⋅;a)=u(t=0,⋅;b)∈C∞(D;R).u_0(\cdot) := u(t=0,\cdot;a)=u(t=0,\cdot;b)\in C^\infty(\mathcal{D};\mathbb{R}).

Then with a given final time T>0T>0,

∥u(t,x;a)−u(t,x;b)∥Lt2Lx2([0,T]×D)≲∥a−b∥L∞(D).\left\|u(t,x;a)-u(t,x;b)\right\|_{L_t^2L_x^2([0,T]\times\mathcal{D})}\lesssim\left\|a-b\right\|_{L^\infty(\mathcal{D})}.

Proof. If not, we can find a sequence by taking ϵ=1k\epsilon=\frac{1}{k} for kk starting at sufficiently large (k≥Kk\ge K),

∃ak,bk∈C∞(D;R+)with∥ak−bk∥L∞≤1k\exists a_k,b_k\in C^\infty(\mathcal{D};\mathbb{R}^+)\quad\text{with}\quad\left\|a_k-b_k\right\|_{L^\infty}\le\frac{1}{k}
s.t.∥u(t,x;ak)−u(t,x;bk)∥Lt2Lx2([0,T]×D)≥1\text{s.t.}\quad\left\|u(t,x;a_k)-u(t,x;b_k)\right\|_{L_t^2L_x^2([0,T]\times\mathcal{D})}\ge1

with the following convergence implied by the Aubin–Lions compactness lemma (Theorem II.5.16 of Boyer and Fabrie [ [3]):

ak→k→∞L∞a0⟹u(⋅,⋅;ak)→k→∞Lt2Lx2([0,T]×D)u(⋅,⋅;a0),a_k\xrightarrow[k\to\infty]{L^\infty}a_0 \qquad\Longrightarrow\qquad u(\cdot,\cdot;a_k)\xrightarrow[k\to\infty]{L_t^2L_x^2([0,T]\times\mathcal{D})}u(\cdot,\cdot;a_0),
bk→k→∞L∞b0=a0⟹u(⋅,⋅;bk)→k→∞Lt2Lx2([0,T]×D)u(⋅,⋅;b0).b_k\xrightarrow[k\to\infty]{L^\infty}b_0=a_0 \qquad\Longrightarrow\qquad u(\cdot,\cdot;b_k)\xrightarrow[k\to\infty]{L_t^2L_x^2([0,T]\times\mathcal{D})}u(\cdot,\cdot;b_0).

Then we have a sequence:

∥u(⋅,⋅;ak)−u(⋅,⋅;bk)∥Lt2Lx2([0,T]×D)≥1∀k≥K.\left\|u(\cdot,\cdot;a_k)-u(\cdot,\cdot;b_k)\right\|_{L_t^2L_x^2([0,T]\times\mathcal{D})}\ge1\qquad\forall k\ge K.

But this contradicts to the uniqueness of the solution to (1). Then we can take an interpolation between aa and bb:

cn=K+1−nK+1a+nK+1bn=1,…,K.c_n=\frac{K+1-n}{K+1}a+\frac{n}{K+1}b\qquad n=1,\ldots,K.

The interpolation satisfies ∥a−c1∥≤1K\left\|a-c_1\right\|\le\frac{1}{K} and ∥cK−b∥≤1K\left\|c_K-b\right\|\le\frac{1}{K}. Then using triangle inequalities with norms on Lt2Lx2([0,T]×D)L_t^2L_x^2([0,T]\times\mathcal{D}) abbreviated,

∥u(⋅,⋅;a)−u(⋅,⋅;b)∥=∥u(⋅,⋅;a)−u(⋅,⋅;c1)+u(⋅,⋅;c1)−⋯+u(⋅,⋅;cK)−u(⋅,⋅;b)∥≤∥u(⋅,⋅;a)−u(⋅,⋅;c1)∥+∥u(⋅,⋅;c1)−u(⋅,⋅;c2)∥+⋯+∥u(⋅,⋅;cK)−u(⋅,⋅;b)∥\begin{aligned} \left\|u(\cdot,\cdot;a)-u(\cdot,\cdot;b)\right\| = \left\|u(\cdot,\cdot;a)-u(\cdot,\cdot;c_1)+u(\cdot,\cdot;c_1)-\cdots+u(\cdot,\cdot;c_K)-u(\cdot,\cdot;b)\right\| \\ &\le\left\|u(\cdot,\cdot;a)-u(\cdot,\cdot;c_1)\right\| +\left\|u(\cdot,\cdot;c_1)-u(\cdot,\cdot;c_2)\right\| +\cdots +\left\|u(\cdot,\cdot;c_K)-u(\cdot,\cdot;b)\right\| \end{aligned}

Each component is bounded by ∥a−c1∥,∥c1−c2∥,…,∥cK−b∥\left\|a-c_1\right\|,\left\|c_1-c_2\right\|,\ldots,\left\|c_K-b\right\|, with norms on L∞(D)L^\infty(\mathcal{D}) abbreviated. □\square

Remark. m>0m>0 is a critical condition. Otherwise, the temperature will never change from u0u_0 where the conductivity is zero, and will eventually change even if conductivity is slightly larger than zero.

Remark. The Lipschitz dependence of the solution on the conductivity ensures numerical stability: small perturbations in the non-degenerate conductivity will result in only minor changes to the temperature dynamics.

1D (special case)

We semi-discretize u(t,x)u(t,x), a(x)a(x), f(t,x)f(t,x) using the finite difference method (FDM) which is first-order accuracy in space, to solve the forward PDE problem:

uj(t):=u(t,(j−1)Δx),aj:=a((j−1)Δx),fj(t):=f(t,(j−1)Δx)u_j(t) := u(t,(j-1)\Delta x), \qquad a_j := a((j-1)\Delta x), \qquad f_j(t) := f(t,(j-1)\Delta x)

for j=1,2,⋯ ,Jj=1,2,\cdots,J. Here, semi-discretize means the discretization is in space xx only, not in time tt. The discretization of the circle is a standard procedure where the x1x_1 and xJx_J are neighbors of each other (see Figure 1a).

Figure 1. Parametrization of D\mathcal{D} for the finite difference method

A finite difference scheme of ∂x(a(x)ux)\partial_x(a(x)u_x) in energy-conservative form is:

∂x(a(x)ux)≈Dx−(ajDx+uj)=Dx−(aj⋅uj+1−ujΔx)=1Δx(aj⋅uj+1−ujΔx−aj−1⋅uj−uj−1Δx)=aj(uj+1−uj)−aj−1(uj−uj−1)Δx2.\begin{aligned} \partial_x(a(x)u_x) &\approx D_x^-(a_jD_x^+u_j)=D_x^-\left(a_j\cdot\frac{u_{j+1}-u_j}{\Delta x}\right)\\ &=\frac{1}{\Delta x}\left(a_j\cdot\frac{u_{j+1}-u_j}{\Delta x}-a_{j-1}\cdot\frac{u_j-u_{j-1}}{\Delta x}\right)\\ &=\frac{a_j(u_{j+1}-u_j)-a_{j-1}(u_j-u_{j-1})}{\Delta x^2}. \end{aligned}

Then the PDE becomes a system of ODEs:

u˙j(t)=1Δx2[ajuj+1(t)−(aj+aj−1)uj(t)+aj−1uj−1(t)]+fj(t),j=1,⋯ ,J.\dot{u}_j(t)=\frac{1}{\Delta x^2}\left[a_ju_{j+1}(t)-(a_j+a_{j-1})u_j(t)+a_{j-1}u_{j-1}(t)\right]+f_j(t),\qquad j=1,\cdots,J.

We can write the system in matrix form:

u˙(t)=A(a)u(t)+f(t),(2)\dot{\mathbf{u}}(t)=A(a)\mathbf{u}(t)+\mathbf{f}(t), \tag*{(2)}

where

u(t)=[u1u2⋯uJ−1uJ]T(t),f(t)=[f1f2⋯fJ−1fJ]T(t)\mathbf{u}(t)=\left[u_1\quad u_2\quad\cdots\quad u_{J-1}\quad u_J\right]^{\mathsf{T}}(t),\qquad \mathbf{f}(t)=\left[f_1\quad f_2\quad\cdots\quad f_{J-1}\quad f_J\right]^{\mathsf{T}}(t)
A(a)=1Δx2[−(a1+aJ)a1aJa1−(a2+a1)a2⋱⋱⋱aJ−2−(aJ−1+aJ−2)aJ−1aJaJ−1−(aJ+aJ−1)]J×JA(a)=\frac{1}{\Delta x^2} \begin{bmatrix} -(a_1+a_J) & a_1 & & & a_J\\ a_1 & -(a_2+a_1) & a_2 & &\\ & \ddots& \ddots& \ddots&\\ & & a_{J-2} & -(a_{J-1}+a_{J-2}) & a_{J-1}\\ a_J & & & a_{J-1} & -(a_J+a_{J-1}) \end{bmatrix}_{J\times J}

By Duhamel’s formula, the solution can be written as

u(t;a)=exp⁡(tA(a))u(0)+∫0texp⁡((t−τ)A(a))f(τ) dτ,(3)\mathbf{u}(t;a)=\exp(tA(a))\mathbf{u}(0)+\int_{0}^{t}\exp((t-\tau)A(a))\mathbf{f}(\tau)\,\mathrm{d}\tau, \tag*{(3)}

where u(0)=[u(t=0,x1)u(t=0,x2)⋯u(t=0,xJ−1)u(t=0,xJ)]T\mathbf{u}(0)=\left[u(t=0,x_1)\quad u(t=0,x_2)\quad\cdots\quad u(t=0,x_{J-1})\quad u(t=0,x_J)\right]^{\mathsf{T}}. This discrete initial condition can be obtained from the given initial condition evaluated on the discretized grid. We split the variables and parameters of a function using semicolon so that u(t;a)\mathbf{u}(t;a) treats tt as the variable and conductivity aa as parameters.

The motion of the sensor is parametrized by s:R+→Ts:\mathbb{R}^{+}\to\mathbb{T}, and the time indices it records temperature is tmt_m for m=1,⋯ ,Mm=1,\cdots,M, with convention that

0<t1<t2<⋯<tM.0<t_1<t_2<\cdots<t_M.

The recorded temperature at each time tmt_m is u(tm,s(tm))u(t_m,s(t_m)). Then we propose the following least square minimizer as the reconstructed conductivity:

a∗=arg min⁡a∈L+2(T)∑m=1M[u(tm,s(tm))−Extract⁡tm,s(tm)(u~(⋅,⋅;a))]2,(4)a^*=\underset{a\in L^2_+(\mathbb{T})}{\operatorname{arg\,min}}\sum_{m=1}^{M}\left[u(t_m,s(t_m))-\operatorname{Extract}_{t_m,s(t_m)}(\widetilde{u}(\cdot,\cdot;a))\right]^2, \tag*{(4)}

where u~(⋅,⋅;a)\widetilde{u}(\cdot,\cdot;a) is the simulated temperature dynamics on (t,x)∈[0,tM]×T(t,x)\in[0,t_M]\times\mathbb{T} using the candidate conductivity aa. We use the nearest neighbor on the grid to represent s(tm)s(t_m),

jm:=arg min⁡j∈{1,2,⋯ ,J}∣xj−s(tm)∣⟹xjm≈s(tm),m=1,2,⋯ ,M.j_m:=\underset{j\in\{1,2,\cdots,J\}}{\operatorname{arg\,min}}\left|x_j-s(t_m)\right|\Longrightarrow x_{j_m}\approx s(t_m),\qquad m=1,2,\cdots,M.

Let ej\mathbf{e}_j denote the jthj^{\text{th}} standard basis for RJ\mathbb{R}^{J} (one-hot vector with 1 in jthj^{\text{th}} coordinate), then we can explicitly define Extract⁡tm,s(tm)(u~(⋅,⋅;a)):=ejmTu(tm;a)\operatorname{Extract}_{t_m,s(t_m)}(\widetilde{u}(\cdot,\cdot;a)):=\mathbf{e}_{j_m}^{\mathsf{T}}\mathbf{u}(t_m;a).

Remark. Although the formulation above is expressed as a least-square minimization, it is crucial to recognize that reconstructing the conductivity aa (or a1,⋯ ,aJa_1,\cdots,a_J after discretization) is not a linear least-square problem. This distinction arises because the governing PDE (1) is linear with respect to the solution uu, but it is nonlinear with respect to the parameter aa, which can be directly seen from (3). Consequently, standard linear least-square techniques are not directly applicable, and specialized numerical methods must be employed to solve for the optimal a∗a^*.

Sensitivity analysis

Based on the spectral analysis of A(a)A(a)1, measurements taken at the same location over a later time become less sensitive to reconstruction errors, given that the external heat source is spatially constant. Formally, we propose the following:

Theorem 2.3. Let the eigenvalue decomposition of A(a)A(a) in (2) (which is Hermitian) be

A=1Δx2QΛQT,whereQ=[v1v2⋯vJ−1vJ]J×J,A=\frac{1}{\Delta x^2}Q\Lambda Q^{\mathsf{T}},\qquad\text{where}\quad Q=[v_1\quad v_2\quad\cdots\quad v_{J-1}\quad v_J]_{J\times J},

and

Λ=diag⁡(λ1,⋯ ,λJ),\Lambda=\operatorname{diag}(\lambda_1,\cdots,\lambda_J),

with the eigenvalues sorted in ascending order:

λ1≤λ2≤⋯≤λJ−1<λJ=0.\lambda_1 \le\lambda_2 \le\cdots\le\lambda_{J-1} < \lambda_J = 0.

Suppose the external heat source is spatially constant, i.e., f(t,x)=f(t)f(t,x)=f(t) for all (t,x)∈R+×T(t,x) \in\mathbb{R}^{+} \times\mathbb{T}. If two measurements, u(tm,s(tm))u(t_m,s(t_m)) and u(tn,s(tn))u(t_n,s(t_n)), where $1 \le m < n \le M$, are taken at the same position but at different times, namely

tn>tm>Δx2−λk,s(tn)=s(tm),for some k=1,2,…,J−1,t_n > t_m > \frac{\Delta x^2}{\sqrt{-\lambda_k}}, \qquad s(t_n)=s(t_m), \qquad\text{for some } k=1,2,\ldots,J-1,

then the measurement u(tm,s(tm))u(t_m,s(t_m)) is more sensitive to perturbations in λk\lambda_k than u(tn,s(tn))u(t_n,s(t_n)).

Proof. Notice that the row sums of AA are zero2, which gives the dominant eigen-pair λJ=0\lambda_J=0, vJ=(1/J)1J\boldsymbol{v}_J=(1/\sqrt{J})\boldsymbol{1}_J, where 1J∈RJ\boldsymbol{1}_J \in\mathbb{R}^J denotes the all-one vector. This eigen-pair corresponds to the constant temperature being a solution to the spatially constant external heat source version of (1) with constant initial data. Hence, (3) becomes

u(t;a)=exp⁡(tA)u(0)+∫0tet−τf(τ) dτ⋅1J.\boldsymbol{u}(t;a)=\exp(tA)\boldsymbol{u}(0)+\int_0^t e^{t-\tau}f(\tau)\,\mathrm{d}\tau\cdot\boldsymbol{1}_J.

Using spectral decomposition on exp⁡(tmA)\exp(t_mA),

exp⁡(tmA)=Qdiag⁡(exp⁡[tmΔx2λ1],…,exp⁡[tmΔx2λJ])QT=∑j=1Jexp⁡(tmΔx2λj)vjvjTu(0).\exp(t_mA)=Q\operatorname{diag}\left(\exp\left[\frac{t_m}{\Delta x^2}\lambda_1\right],\ldots,\exp\left[\frac{t_m}{\Delta x^2}\lambda_J\right]\right)Q^{\mathsf{T}}=\sum_{j=1}^{J}\exp\left(\frac{t_m}{\Delta x^2}\lambda_j\right)\boldsymbol{v}_j\boldsymbol{v}_j^{\mathsf{T}}\boldsymbol{u}(0).

Measurement u(tm,s(tm))u(t_m,s(t_m)) can be extracted from u(tm;a)\boldsymbol{u}(t_m;a) by the dot product with ejm\boldsymbol{e}_{j_m}:

u(tm,s(tm))=ejmT∑j=1Jexp⁡(tmΔx2λj)vjvjTu(0)+∫0tmetm−τf(τ) dτ.u(t_m,s(t_m))=\boldsymbol{e}_{j_m}^{\mathsf{T}}\sum_{j=1}^{J}\exp\left(\frac{t_m}{\Delta x^2}\lambda_j\right)\boldsymbol{v}_j\boldsymbol{v}_j^{\mathsf{T}}\boldsymbol{u}(0)+\int_0^{t_m}e^{t_m-\tau}f(\tau)\,\mathrm{d}\tau.

We take derivatives with respect to λk\lambda_k to analyze the sensitivity:

∂u(tm,s(tm))∂λk=∂∂λkexp⁡(tmΔx2λk)ejmTvkvkTu(0)=tmΔx2exp⁡(tmΔx2λk)ejmTvkvkTu(0).\begin{aligned} \frac{\partial u(t_m,s(t_m))}{\partial\lambda_k} =\frac{\partial}{\partial\lambda_k}\exp\left(\frac{t_m}{\Delta x^2}\lambda_k\right)\boldsymbol{e}_{j_m}^{\mathsf{T}}\boldsymbol{v}_k\boldsymbol{v}_k^{\mathsf{T}}\boldsymbol{u}(0) \\ &=\frac{t_m}{\Delta x^2}\exp\left(\frac{t_m}{\Delta x^2}\lambda_k\right)\boldsymbol{e}_{j_m}^{\mathsf{T}}\boldsymbol{v}_k\boldsymbol{v}_k^{\mathsf{T}}\boldsymbol{u}(0). \end{aligned}

Similarly,

∂u(tn,s(tn))∂λk=tnΔx2exp⁡(tnΔx2λk)ejnTvkvkTu(0),\frac{\partial u(t_n,s(t_n))}{\partial\lambda_k} =\frac{t_n}{\Delta x^2}\exp\left(\frac{t_n}{\Delta x^2}\lambda_k\right)\boldsymbol{e}_{j_n}^{\mathsf{T}}\boldsymbol{v}_k\boldsymbol{v}_k^{\mathsf{T}}\boldsymbol{u}(0),

and since the position extraction vectors are the same, i.e., ejm=ejn\boldsymbol{e}_{j_m}=\boldsymbol{e}_{j_n}, by s(tm)=s(tn)s(t_m)=s(t_n), it suffices to compare the coefficients:

tmΔx2exp⁡(tmΔx2λk)>tnΔx2exp⁡(tnΔx2λk),\frac{t_m}{\Delta x^2}\exp\left(\frac{t_m}{\Delta x^2}\lambda_k\right)>\frac{t_n}{\Delta x^2}\exp\left(\frac{t_n}{\Delta x^2}\lambda_k\right),

which is clear from the monotonically decreasing region of the rescaled function x↦xexp⁡(λx)x \mapsto x\exp(\lambda x):

ddt[tΔx2exp⁡(tΔx2λk)]=1Δx2(1+λkt2Δx2)exp⁡(tΔx2λk)<0,∀t>Δx2−λk.\frac{\mathrm{d}}{\mathrm{d}t}\left[\frac{t}{\Delta x^{2}}\exp\left(\frac{t}{\Delta x^{2}}\lambda_k\right)\right]=\frac{1}{\Delta x^{2}}\left(1+\lambda_k\frac{t^{2}}{\Delta x^{2}}\right)\exp\left(\frac{t}{\Delta x^{2}}\lambda_k\right)<0,\quad\forall t>\frac{\Delta x^{2}}{\sqrt{-\lambda_k}}.

Remark. From the calculation above, we also conclude that when tm≫1t_m\gg1,

u(tm,s(tm))−1J1Tu(0)⏟average−∫0tmetm−τf(τ) dτ⏟external heat=∑j=1J−1exp⁡(tmΔx2λj)ejmTvjvjTu(0)⏟≈0.u(t_m,s(t_m))-\underbrace{\frac{1}{J}\mathbf{1}^{\mathsf{T}}u(0)}_{\text{average}}-\underbrace{\int_{0}^{t_m}e^{t_m-\tau}f(\tau)\,\mathrm{d}\tau}_{\text{external heat}}=\underbrace{\sum_{j=1}^{J-1}\exp\left(\frac{t_m}{\Delta x^{2}}\lambda_j\right)\mathbf{e}_{j m}^{\mathsf{T}}\mathbf{v}_j\mathbf{v}_j^{\mathsf{T}}u(0)}_{\approx0}.

Remark. The sensitivity can be understood as follows: if there is a small perturbation in the reconstructed conductivity from the true conductivity aa (in terms of the eigenvalues of A(a)A(a)), then the calculated temperature based on the perturbed conductivity will significantly deviate from the actual measurement. We can see from the proof that the sensitivity with respect to time goes up and down and acts like the function t↦texp⁡(λkt)t \mapsto t\exp(\lambda_k t). This implies that if the measurement is taken too early, the temperature is primarily influenced by the initial condition. Conversely, if the measurement is taken too late, the temperature is dominated by the average temperature and the accumulated effect of the external heat source (i.e., equilibrium). In both cases, the temperature is mainly driven by the initial condition, which is independent of the conductivity. Therefore, there exists a time interval during which the temperature measurement is highly sensitive to errors in the reconstruction, making this period crucial for accurate conductivity estimation. Moreover, this interval is shorter for smaller λk\lambda_k. In practice, if the measurements are taken too early for λk\lambda_k’s, we can always continue recording until the system enters the highly sensitive interval. However, if the measurement is taken too late for λJ−1\lambda_{J-1}, then it’s also too late for all other smaller λk\lambda_k’s, which can be treated as garbage data, and the reconstruction of conductivity is inevitably prone to large errors.

Extension to 2D space

Consider (1) on T2=R2/Z2∼[0,1)2\mathbb{T}^{2}=\mathbb{R}^{2}/\mathbb{Z}^{2}\sim[0,1)^{2} with x=(x,y)\mathbf{x}=(x,y):

ut(t,x,y)=∂x[a(x,y)∂xu]+∂y[a(x,y)∂yu]+f(t,x,y).(5)u_t(t,x,y)=\partial_x[a(x,y)\partial_xu]+\partial_y[a(x,y)\partial_yu]+f(t,x,y). \tag*{(5)}

Again, u(t,x,y)u(t,x,y) is the temperature at time tt and spatial position (x,y)(x,y) and a(x,y)a(x,y) is the surface thermal conductivity. We inherit the 1D convention that xi=(i−1)Δxx_i=(i-1)\Delta x and yj=(j−1)Δyy_j=(j-1)\Delta y for i∈{1,2,⋯ ,K}i\in\{1,2,\cdots,K\} and j∈{1,2,⋯ ,J}j\in\{1,2,\cdots,J\}, where KΔx=1K\Delta x=1 and JΔy=1J\Delta y=1 (so that they are grid points on a uniform grid on the unit square). Similar as 1D case, we semi-discretize the spatial domain according to fig. 1b into

ui,j(t):=u(t,xi,yj),ai,j:=a(xi,yj),fi,j(t):=f(t,xi,yj).u_{i,j}(t):=u(t,x_i,y_j),\qquad a_{i,j}:=a(x_i,y_j),\qquad f_{i,j}(t):=f(t,x_i,y_j).

We will discretize the right-hand side of (5) as follows 3:

Dx−(ai,jDx+ui,j)+Dy−(ai,jDy+ui,j)≈1Δx[ai,j⋅ui+1,j−ui,jΔx−ai−1,j⋅ui,j−ui−1,jΔx]+1Δy[ai,j⋅ui,j+1−ui,jΔy−ai,j−1⋅ui,j−ui,j−1Δy].D_x^{-}\left(a_{i,j}D_x^{+}u_{i,j}\right)+D_y^{-}\left(a_{i,j}D_y^{+}u_{i,j}\right)\approx\frac{1}{\Delta x}\left[a_{i,j}\cdot\frac{u_{i+1,j}-u_{i,j}}{\Delta x}-a_{i-1,j}\cdot\frac{u_{i,j}-u_{i-1,j}}{\Delta x}\right] +\frac{1}{\Delta y}\left[a_{i,j}\cdot\frac{u_{i,j+1}-u_{i,j}}{\Delta y}-a_{i,j-1}\cdot\frac{u_{i,j}-u_{i,j-1}}{\Delta y}\right].

For simplicity, we take Δx=Δy\Delta x = \Delta y, which implies K=JK = J, leading to

∇⋅[a(xi,yj)∇u(t,xi,yj)]≈1Δx2[ai,j(ui+1,j−ui,j)−ai−1,j(ui,j−ui−1,j)+ai,j(ui,j+1−ui,j)−ai,j−1(ui,j−ui,j−1)].\nabla\cdot[a(x_i,y_j)\nabla u(t,x_i,y_j)] \approx\frac{1}{\Delta x^2}\left[a_{i,j}(u_{i+1,j}-u_{i,j})-a_{i-1,j}(u_{i,j}-u_{i-1,j})+a_{i,j}(u_{i,j+1}-u_{i,j})-a_{i,j-1}(u_{i,j}-u_{i,j-1})\right].

Then we have a system of ODEs (with i,j∈{1,2,⋯ ,J}i,j \in\{1,2,\cdots,J\}):

u˙i,j=1Δx2[ai,jui+1,j+ai−1,jui−1,j+ai,jui,j+1+ai,j−1ui,j−1−(2ai,j+ai−1,j+ai,j−1)ui,j]+fi,j.(6)\dot{u}_{i,j}=\frac{1}{\Delta x^2}\left[a_{i,j}u_{i+1,j}+a_{i-1,j}u_{i-1,j}+a_{i,j}u_{i,j+1}+a_{i,j-1}u_{i,j-1}-(2a_{i,j}+a_{i-1,j}+a_{i,j-1})u_{i,j}\right]+f_{i,j}. \tag*{(6)}

This system is of the form of (2) where A(a)∈RJ2×J2A(a) \in\mathbb{R}^{J^2 \times J^2}. The unique solution can again be obtained by Duhamel’s formula, and we denote it as u(⋅,⋅;a)u(\cdot,\cdot;a). Then we consider the observation of a sensor whose position at time tt is s(t)∈T2s(t) \in\mathbb{T}^2. Therefore, the recorded temperature by this sensor is {u(tm,s(tm))}m=1,⋯ ,M\{u(t_m,s(t_m))\}_{m=1,\cdots,M}. Similar as 1D, we need the conductivity that minimizes the difference between the recorded temperature and simulated temperature based on the candidate conductivity aa, so we use exactly the same minimizer of (4), where the Extract operation is also based on nearest neighbor approximation, which implies s(tm)≈(xim,yjm)s(t_m) \approx(x_{i_m},y_{j_m}).

Sensor Motion, Initial Condition, and External Heat Source

Other than the temperature uu and conductivity aa, the terms in the PDE need to be considered are:

  • Sensor motion: s(t)s(t);

  • Initial condition u(t=0,⋅)u(t=0,\cdot);

  • External heat source f(t,⋅)f(t,\cdot).

Design of Sensor Motion

The sensitivity analysis above provides insight into the advantages of using moving sensors over static ones. While static sensors are limited to detecting reconstruction errors in conductivity within their immediate vicinity, moving sensors are able to detect errors across the entire region they traverse. As mentioned in Remark 2.3, there is a specific time interval during which temperature measurements are highly sensitive to reconstruction errors in conductivity. During this critical period, the more of the spatial domain the sensors traverse, the better the reconstruction quality. To fully exploit this sensitivity window, it is advantageous for moving sensors to cover the entire spatial domain. This approach ensures that sensors gather comprehensive data from across the domain during the most sensitive phase, thereby improving the precision and accuracy of the reconstruction.

Sensor Placement and Movements for D=T\mathcal{D} = \mathbb{T}

In one-dimensional settings, a single sensor can effectively traverse the entire domain by moving along T\mathbb{T} at a constant speed over the unit time interval t∈[0,1]t \in[0,1] (by setting tM=1t_M=1). This approach ensures ergodicity, as the sensor returns to its initial position after covering the entire domain, thereby maximizing its ability to capture detailed temperature data within the given timeframe.

Sensor Placement and Movements for D=T2\mathcal{D} = \mathbb{T}^2

Extending the methodology of using a single sensor to traverse the domain to higher dimensions presents significant challenges. In dimensions greater than one, smooth curves occupy zero measure [15], meaning that a sensor following such a trajectory would collect temperature data from an infinitesimally small portion of the domain. Attempts to traverse space-filling curves (or their finite fractal levels) within the high-sensitivity window leads to infinite (or large) moving speeds of the sensor, which leads to physical infeasibility. Consequently, the data gathered would be insufficient for accurately reconstructing the conductivity across the entire spatial region, limiting the reconstruction accuracy to only a localized area.

Given these constraints, the physical movement of sensors must prioritize continuity and regularity, often restricting trajectories to piecewise constant functions with bounded jumps after spatial discretization. This means that the sensor either moves between adjacent pixels or stays in place, i.e.,

∣s(tm)−s(tm−1)∣Manhattan on discretized T2≤Δx,m=1,⋯ ,M.\left|s(t_m)-s(t_{m-1})\right|_{\mathrm{Manhattan\ on\ discretized}\ \mathbb{T}^2} \leq\Delta x,\qquad m=1,\cdots,M.

This limitation makes it impossible for a single sensor to cover the entire discretized domain T2\mathbb{T}^2 within a reasonable timeframe. Furthermore, temperature measurements taken at later times tend to contribute less effectively to the reconstruction of the conductivity.

To overcome these limitations, deploying multiple temperature sensors becomes essential. By distributing sensors across different locations and ensuring their trajectories cover reasonably large regions of the domain, broader spatial coverage is achieved early in the measurement process. This strategic deployment enhances the accuracy of thermal conductivity reconstruction and mitigates the inefficiencies associated with single-sensor approaches. Therefore, utilizing multiple sensors with well-planned trajectories is crucial for comprehensive and reliable data collection in higher-dimensional domains.

As shown in Figure 2, we place four sensors at the corners of four concentric squares, each sensor moving along its respective orbit. These squares are referred to as the orbits of the sensors, as they always move along the sides of their initial paths. The side lengths of these four squares are 15\frac{1}{5}, 25\frac{2}{5}, 35\frac{3}{5}, and 45\frac{4}{5}, from the innermost to the outermost square, all centered at (x=12,y=12)(x=\frac{1}{2},y=\frac{1}{2}). Each sensor starts from a corner of its orbit, and the initial positions are shown as ∙\bullet, ∗\ast, ■\blacksquare and ∘\circ, respectively, along with the contours of the initial temperature.

Image not extracted

Figure 2. Initial positions s1(0),⋯ ,s4(0)s_1(0),\cdots,s_4(0) and trajectories of the four sensors on T2\mathbb{T}^2

The sensors travel counterclockwise along their orbits to the next corner, continuing this pattern until t=1t=1. At t=1t=1, they have completed 8, 4, 2, and 1 cycles along their orbits, respectively, from the innermost to the outermost square, all moving at constant speeds. By this design, we ensure that every point in T2\mathbb{T}^2 will be approached by at least one sensor during its high-sensitivity window.

Initial Condition and External Heat Source

We also consider the frequency domain for the 1D spatial domain, where we take x∈[0,1)x \in[0,1) from the representative interval of T\mathbb{T}:

u(t,x)=∑n∈Zu^n(t)e2πinx,u^n(t)=∫01u(t,x)e−2πinx dx,u^(t)∈ℓ2(Z)∀t∈[0,∞),u(t,x)=\sum_{n\in\mathbb{Z}}\widehat{u}_n(t)e^{2\pi inx},\qquad\widehat{u}_n(t)=\int_0^1 u(t,x)e^{-2\pi inx}\,\mathrm{d}x,\qquad\widehat{u}(t)\in\ell^2(\mathbb{Z})\quad\forall t\in[0,\infty),
f(t,x)=∑n∈Zf^n(t)e2πinx,f^n(t)=∫01f(t,x)e−2πinx dx,f^(t)∈ℓ2(Z)∀t∈[0,∞),f(t,x)=\sum_{n\in\mathbb{Z}}\widehat{f}_n(t)e^{2\pi inx},\qquad\widehat{f}_n(t)=\int_0^1 f(t,x)e^{-2\pi inx}\,\mathrm{d}x,\qquad\widehat{f}(t)\in\ell^2(\mathbb{Z})\quad\forall t\in[0,\infty),
a(x)=∑n∈Za^(n)e2πinx,a^(n)=∫01a(x)e−2πinx dx,a^∈ℓ2(Z).a(x)=\sum_{n\in\mathbb{Z}}\widehat{a}(n)e^{2\pi inx},\qquad\widehat{a}(n)=\int_0^1 a(x)e^{-2\pi inx}\,\mathrm{d}x,\qquad\widehat{a}\in\ell^2(\mathbb{Z}).

Here, u^(t)\widehat{u}(t), f^(t)\widehat{f}(t), and a^\widehat{a} denote the Fourier coefficients of u(t,⋅)u(t,\cdot), f(t,⋅)f(t,\cdot), and aa respectively. Substituting these expansions into Eq. (1) transforms it into an infinite system of ODEs:

u^˙n(t)=−4π2n∑m∈Zma^(n−m)u^m(t)+f^n(t),∀n∈Z.\dot{\widehat{u}}_n(t)=-4\pi^2n\sum_{m\in\mathbb{Z}}m\widehat{a}(n-m)\widehat{u}_m(t)+\widehat{f}_n(t),\qquad\forall n\in\mathbb{Z}.

This can be written in matrix form:

u^˙(t)=−4π2Da^(Z)Du^(t)+f^(t),(7)\dot{\widehat{u}}(t)=-4\pi^2D\widehat{a}(Z)D\widehat{u}(t)+\widehat{f}(t), \tag*{(7)}

where a^(Z)\widehat{a}(Z) is a bi-infinite Toeplitz matrix4{}^{4} whose entries are given by a^(n−m)\widehat{a}(n-m) for each pair (n,m)∈Z(n,m)\in\mathbb{Z}. The diagonal matrix DD and the bi-infinite Toeplitz matrix ZZ are defined as

D=[⋱−2−1012⋱]Z×Z,Z=[⋱⋮⋮⋮⋮⋮⋱⋯0−1−2−3−4⋯⋯10−1−2−3⋯⋯210−1−2⋯⋯3210−1⋯⋯43210⋯⋱⋮⋮⋮⋮⋮⋱]Z×Z.D=\left[\begin{array}{ccccc} \ddots& & & & \\ & -2 & & & \\ & & -1 & & \\ & & & 0 & \\ & & & & 1 \\ & & & & & 2 \\ & & & & & & \ddots \end{array}\right]_{\mathbb{Z}\times\mathbb{Z}}, \qquad Z=\left[\begin{array}{ccccccc} \ddots& \vdots& \vdots& \vdots& \vdots& \vdots& \ddots\\ \cdots& 0 & -1 & -2 & -3 & -4 & \cdots\\ \cdots& 1 & 0 & -1 & -2 & -3 & \cdots\\ \cdots& 2 & 1 & 0 & -1 & -2 & \cdots\\ \cdots& 3 & 2 & 1 & 0 & -1 & \cdots\\ \cdots& 4 & 3 & 2 & 1 & 0 & \cdots\\ \ddots& \vdots& \vdots& \vdots& \vdots& \vdots& \ddots \end{array}\right]_{\mathbb{Z}\times\mathbb{Z}}.

Again, by Duhamel’s formula,

u^(t;a)=exp⁡[−4π2tDa^(Z)D]u^(0)+∫0texp⁡[−4π2(t−τ)Da^(Z)D]f^(τ) dτ.(8)\widehat{u}(t;a)=\exp\left[-4\pi^2tD\widehat{a}(Z)D\right]\widehat{u}(0)+\int_0^t\exp\left[-4\pi^2(t-\tau)D\widehat{a}(Z)D\right]\widehat{f}(\tau)\,\mathrm{d}\tau. \tag*{(8)}

Notice that ker⁡(D)=span⁡{e0}\ker(D)=\operatorname{span}\{\mathbf{e}_0\}, which again confirms Proposition 2.1. Since we need at least one of the initial condition or the external heat source to be non-constant in space, we choose the initial condition to be non-constant while keeping the external heat source constant in space, i.e., f(t)=f(t,x)f(t)=f(t,x) on (t,x)∈[0,∞)×T(t,x)\in[0,\infty)\times\mathbb{T}. This simplifies the problem to:

u^(t;a)=exp⁡[−4π2tDa^(Z)D]u^(0)+e0∫0te−4π2(t−τ)f(τ) dτ.\widehat{u}(t;a)=\exp\left[-4\pi^{2}tD\widehat{a}(Z)D\right]\widehat{u}(0)+e_{0}\int_{0}^{t}e^{-4\pi^{2}(t-\tau)}f(\tau)\,\mathrm{d}\tau.

As for the initial condition, we choose the lowest frequency based on the Fourier coefficients:

u^(0)=−i2(e1−e−1)⟺u(t=0,x)=sin⁡(2πx),x∈T.\widehat{u}(0)=-\frac{i}{2}(e_{1}-e_{-1})\Longleftrightarrow u(t=0,x)=\sin(2\pi x),\qquad x\in\mathbb{T}.

In the 2D problem, we adopt the same criteria:

u(t=0,x,y)=cos⁡(2πx)cos⁡(2πy),(x,y)∈T2.(9)u(t=0,x,y)=\cos(2\pi x)\cos(2\pi y),\qquad(x,y)\in\mathbb{T}^{2}. \tag*{(9)}

As long as the external heat source is spatially constant, it neither improves nor worsens the resolution of the reconstruction. Therefore, we choose the following form for the external heat source:

f(t,x)=sin⁡(πt),(t,x)∈[0,∞)×T;(10)f(t,x)=\sin(\pi t),\qquad(t,x)\in[0,\infty)\times\mathbb{T}; \tag*{(10)}
f(t,x,y)=sin⁡(2πt),(t,x,y)∈[0,∞)×T2.(11)f(t,x,y)=\sin(2\pi t),\qquad(t,x,y)\in[0,\infty)\times\mathbb{T}^{2}. \tag*{(11)}

Overcoming Sparsity in Temperature Measurements

Reduction of Degrees of Freedom (DoF) for D=TD=\mathbb{T}

The inverse heat equation poses significant challenges due to the non-convexity of the loss function (4), rendering analytical minimization approaches often infeasible. Additionally, the number of temperature measurements MM is typically less than the desired spatial resolution JJ (M<JM<J), leading to over-parameterization if all degrees of freedom are considered. To address this, we restrict our search to low-dimensional subspaces of L2(T)L^{2}(\mathbb{T}) by parameterizing the logarithm of the conductivity a(⋅)a(\cdot) using Fourier series:

ln⁡(a(x;θ))=θ0+∑k=1⌊dim⁡θ/2⌋θ2k−1sin⁡(2kπx)+∑k=1⌊dim⁡θ/2⌋θ2kcos⁡(2kπx),\ln(a(x;\theta))=\theta_{0}+\sum_{k=1}^{\lfloor\dim\theta/2\rfloor}\theta_{2k-1}\sin(2k\pi x)+\sum_{k=1}^{\lfloor\dim\theta/2\rfloor}\theta_{2k}\cos(2k\pi x),

which is inspired from the bi-infinite Toeplitz matrix a^(Z)\widehat{a}(Z) in (8). This parameterization allows flexibility in dim⁡θ\dim\theta, constrained by an upper bound to reflect the typical regularity of a(⋅)a(\cdot) or the presence of only a few abrupt transitions corresponding to material changes. This flexibility allows us to avoid over-parameterization by setting dim⁡θ≪M\dim\theta\ll M, even when the number of temperature measurements is limited. Consequently, we refer to θ\theta as the model parameter, as it governs the reduced-dimensional representation of the conductivity, ensuring flexibility and avoiding over-parameterization while maintaining the positivity of a(⋅)a(\cdot) without additional constraints, in line with the conventions in Stuart [ [16] and Beskos et al. [ [1].

In the machine learning framework, the loss function for a candidate θ\theta can be derived from (4) as

L(θ):=1M∑m=1M[u(tm,s(tm))−ejmTu(tm;a(⋅;θ))]2,(12)L(\theta):=\frac{1}{M}\sum_{m=1}^{M}\left[u(t_m,s(t_m))-e_{jm}^{\mathsf{T}}\mathbf{u}(t_m;a(\cdot;\theta))\right]^2, \tag*{(12)}

which is the standard mean-square error (MSE) between the measured temperature and calculated temperature based on reconstructed conductivity.

Notice that incrementally increasing the number of Fourier modes in the search for ln⁡(a)\ln(a)’s coefficients corresponds to retaining additional diagonals in the bi-infinite Toeplitz matrix a^(Z)\hat{a}(Z) as defined in (8). Specifically, a search dimension of dim⁡θ=2k+1\dim\theta= 2k + 1 includes the main diagonal and its kk nearest diagonals on either side, while truncating all other diagonals to zero. For example:

  • dim⁡θ=1\dim\theta= 1: Retains only the main diagonal a^(0)\hat{a}(0).

  • dim⁡θ=3\dim\theta= 3: Incorporates the main diagonal and the two adjacent diagonals a^(0)\hat{a}(0), a^(−1)\hat{a}(-1), a^(1)\hat{a}(1).

  • dim⁡θ=5\dim\theta= 5: Includes four nearest diagonals around the main diagonal, a^(0)\hat{a}(0), a^(−1)\hat{a}(-1), a^(1)\hat{a}(1), a^(−2)\hat{a}(-2), a^(2)\hat{a}(2).

This systematic inclusion of diagonals allows for progressively capturing more complex features of conductivity, enhancing the reconstruction accuracy as the search dimension increases. Thus, we propose to adaptively search the Fourier series coefficients from low frequency to high frequency; if the gradient with respect to the current coefficients nearly vanishes, we then allow higher-frequency coefficients to deviate from zero.

Algorithm 1  Adaptive Fourier Series Gradient Descent
1: procedure AdaptiveFS-GD(N, γ, ϵ, MaxEpoch)
2:     θ ← (0,)                                      ▷ Initial guess θ₀ = 0 with dim θ = 1
3:     for n = 1, . . . , MaxEpoch do
4:         g ← ∇θL(θ)                                ▷ loss.backward()
5:         if ∥g∥² < ϵ and dim(θ) ≤ 2N then
6:             θ ← (θ, 0)                             ▷ Include more Fourier modes
7:         else
8:             θ ← θ − γg                             ▷ Steepest/gradient descent
9:         end if
10:    end for
11:    return θ
12: end procedure

Multiple Sensors for D=T2\mathcal{D} = \mathbb{T}^2

In extending our approach to two-dimensional domains, the challenge of insufficient temperature measurements is mitigated by increasing the number of sensors, as demonstrated in subsubsection 3.1.2. Consequently, the necessity to reduce DoF as in the one-dimensional case becomes less critical. Therefore, we define the search space for the 2D case as RJ×J≃RJ2\mathbb{R}^{J \times J} \simeq\mathbb{R}^{J^2}, where each parameter corresponds to the logarithm of conductivity at discrete grid points:

θ(i−1)J+j=ln⁡a(xi,yj),i,j∈{1,2,…,J}.\theta_{(i-1)J+j} = \ln a(x_i, y_j), \qquad i,j \in\{1,2,\ldots,J\}.

Here, the dimension dim⁡θ=J2\dim\theta= J^2 is fixed, and the indexing of θ\theta starts from 1 instead of 0, unlike in the 1D case, as there is no even/odd parity of the indices corresponding to cosine and sine basis functions.

Similar as 1D, loss is taken to be the MSE between prediction and observation:

L(θ):=1MK∑k=1K∑m=1M[u(tm,sk(tm))−Extract⁡tm,sk(tm)(u~(⋅,⋅;a(⋅;θ)))]2(13)L(\theta) := \frac{1}{MK} \sum_{k=1}^{K} \sum_{m=1}^{M} \left[u(t_m, s_k(t_m)) - \operatorname{Extract}_{t_m,s_k(t_m)}\left(\tilde{u}(\cdot,\cdot;a(\cdot;\theta))\right)\right]^2 \tag*{(13)}

where KK is the number of sensors and sk(t)s_k(t) is the position of the kthk^{\mathrm{th}} sensor at time tt. We assume that all sensors are synchronized and record temperature measurements simultaneously.

The reason for adopting the natural pixel basis is that a naive reduction of DoF using a 2D Fourier basis does not work well in two dimensions. A similar effect is observed in JPEG compression, where an image is segmented into 8×88 \times8 pixel blocks, and each block is decomposed into its 2D Fourier series coefficients, instead of directly decomposing the full image into Fourier cosine series as in Wallace [ [18]. More adaptive approaches are proposed and tested in Du and Stechmann [ [6], but we do not apply reduction of DoF in 2D due to the difficulty in implementation.

Numerical Experiments

We leverage the automatic differentiation capabilities of PyTorch [11, 12] and the one-step ODE solver odeint provided by the torchdiffeq library [5]. Automatic differentiation enables the computation of ∇θL(θ)\nabla_\theta L(\theta) (via loss.backward() in algorithm 1) by evaluating the loss function only once, thereby requiring only a single forward PDE solve per gradient calculation.

In contrast, finite difference methods for gradient estimation necessitate multiple evaluations of the loss function. Specifically, the gradient component in the direction of ej\mathbf{e}_j is approximated as:

∇θL(θ)⋅ej≈1Δθ[L(θ+Δθej)−L(θ)],j=1,…,dim⁡θ,\nabla_\theta L(\theta) \cdot\mathbf{e}_j \approx\frac{1}{\Delta\theta}\left[L(\theta+\Delta\theta\mathbf{e}_j)-L(\theta)\right], \qquad j=1,\ldots,\dim\theta,

and the full gradient ∇θL(θ)\nabla_\theta L(\theta) requires dim⁡θ+1\dim\theta+1 evaluations of the loss function. This approach significantly increases computational costs, as each additional degree of freedom requires an extra forward PDE solve. Even with parallel computation, the total number of forward PDE solves remains prohibitive.

To assess the quality of conductivity reconstructions, we define the relative error as:

Relative error:=∥areconstructed−atruth∥∥atruth∥.(14)\text{Relative error} := \frac{\lVert a_{\mathrm{reconstructed}}-a_{\mathrm{truth}}\rVert}{\lVert a_{\mathrm{truth}}\rVert}. \tag*{(14)}

For the one-dimensional case, we set J=100J=100, computing the norm in R100\mathbb{R}^{100}. In the two-dimensional case, we choose J=32J=32 and employ the Frobenius norm. Both cases assume the ground truth conductivity lies within the range of ∼10−2\sim10^{-2}, which ensures the time interval [0,1][0,1] primarily overlaps with the high-sensitivity window. If the conductivity is too large (>1>1), the heat dissipates too quickly, and only a small portion of the temperature measurements fall within the high-sensitivity window. In such cases, the available data becomes insufficient for accurate conductivity reconstruction. Similarly, we assume the minimum conductivity is of the same order as the maximum. The scenario where the minimum conductivity is orders of magnitude smaller than the maximum, a typical multi-scale problem, falls outside the scope of this research and will be addressed in future work.

Single sensor on T with discretization J=100

In the numerical experiments, we will investigate M=J=100M=J=100 while N=9N=9 in algorithm 1, leading to dim⁡θ≤19\dim\theta\le19. Remarkably, the code in Supplementary Materials demonstrates how the forward PDE can be efficiently solved with automatic differentiation enabled. This allows for efficient computation of the gradient of the loss function with respect to the input parameter θ\theta via backpropagation [14], eliminating the need for finite difference methods and significantly reducing computational overhead while improving efficiency.

Testcases: Heaviside, PieceLinear3S, PieceLinear4W

We design three different test cases to evaluate the performance of the conductivity reconstruction: Heaviside, PieceLinear3S, and PieceLinear4W.

aHeaviside(x)=1100⋅{1,x∈[0,12)2,x∈[12,1)a^{\mathrm{Heaviside}}(x)=\frac{1}{100}\cdot\begin{cases} 1, & x\in\left[0,\frac{1}{2}\right) \\ 2, & x\in\left[\frac{1}{2},1\right) \end{cases}
aPieceLinear3S(x)=1100⋅{2−3x,x∈[0,13)6x−1,x∈[13,23)5−3x,x∈[23,1)a^{\mathrm{PieceLinear3S}}(x)=\frac{1}{100}\cdot\begin{cases} 2-3x, & x\in\left[0,\frac{1}{3}\right) \\ 6x-1, & x\in\left[\frac{1}{3},\frac{2}{3}\right) \\ 5-3x, & x\in\left[\frac{2}{3},1\right) \end{cases}
aPieceLinear4W(x)=1100⋅{2−4x,x∈[0,14)8x−1,x∈[14,12)7−8x,x∈[12,34)4x−2,x∈[34,1)a^{\mathrm{PieceLinear4W}}(x)=\frac{1}{100}\cdot\begin{cases} 2-4x, & x\in\left[0,\frac{1}{4}\right) \\ 8x-1, & x\in\left[\frac{1}{4},\frac{1}{2}\right) \\ 7-8x, & x\in\left[\frac{1}{2},\frac{3}{4}\right) \\ 4x-2, & x\in\left[\frac{3}{4},1\right) \end{cases}

The latter two are named for their “piecewise linear functions”, with PieceLinear3S having 3 segments resembling a straightened S and PieceLinear4W having 4 segments resembling a W. The blue

circled line in Figure 3a represents the temperature measurement obtained from the sensor (by solving the forward problem), while the red squared line is the calculated temperature measurement based on the initial guess of conductivity (red squared line in Figure 3c). The L2L^2 difference between these two lines in Figure 3a represents the loss, while L2L^2 difference between the two lines in Figure 3c is the error. After training, the calculated temperature measurement is plotted as red squared line in Figure 3b, which closely matches the temperature recorded by the sensor. Additionally, the underlying reconstructed conductivity is plotted as the red squared line in Figure 3d, showing significant improvement over the constant initial guess. The same setup is performed on two other conductivity profiles as shown in Figure 4 and Figure 5, which leads to similar improvements.

Four plots showing initial loss, final loss, initial error, and final error for the Heaviside function

Figure 3. 1D test case: Heaviside function

Image not extracted

Figure 4. 1D test case: Piecewise Linear (3-piece, S shape)

Image not extracted

Figure 5. 1D test case: Piecewise Linear (4-piece, W shape)

Optimization record

We illustrates the progress of algorithm 1 through three key metrics: The number of Fourier basis functions, the loss, and the error over 500 epochs (meaning loss.backward() is called 500 times). fig:6a shows how the number of Fourier basis functions evolves throughout the training

process, and the jumps occur when gradient of the current model parameter ∇θL\nabla_\theta L nearly vanishes, which eventually reaches the cap dim⁡θ≤19\dim\theta\le19. Additionally, fig. 6b and fig. 6c demonstrate the corresponding reduction in the loss and error, respectively, indicating steady improvement in model performance as training progresses. The vertical axis of Figure 6c can be interpreted as the logarithm of eq. (14) since they only differ by a constant.

Image not extracted

Figure 6. Training logs

Moving sensors on T2T^2 with Discretization J=32J = 32

For the 2D problem, we utilize MNIST grayscale images [8] as test cases to evaluate the performance of the conductivity reconstruction using moving sensors. The original MNIST images, which are 28×2828 \times28 pixels with grayscale values in the range [0,1][0,1], are first resized to 32×3232 \times32 (the nearest dyadic number) using torchvision.transforms.Resize((32, 32)) [17]. We then apply an affine transformation to scale the pixel values to the range [1e−2,2e−2][1\mathrm{e}{-2},2\mathrm{e}{-2}], representing conductivity values. This transformation is defined as:

a(xi,yj)=p(i,j)+1100,a(x_i,y_j) = \frac{p(i,j)+1}{100},

where p(i,j)p(i,j) denotes the pixel value at the ii-th row and jj-th column of the grayscale image.

In this transformation:

  • Pixels in the background of the MNIST images (with values near 0) are mapped to a conductivity of approximately 0.01, corresponding to purple regions in Figure 7a.

  • Pixels corresponding to the handwritten digits (with values near 1) are mapped to a conductivity of approximately 0.02, corresponding to yellow regions in Figure 7a.

Image not extracted

Figure 7. MNIST[0] as atrutha_{\mathrm{truth}}, (9) as initial condition and (11) as external heat source

This transformation ensures that the conductivity values fall within a reasonable range, preventing overly fast heat diffusion during the simulation, which could result in insufficient data being collected within each sensor’s high-sensitivity window.

The moving sensors are then deployed to capture temperature data from the forward PDE solver using conductivity as these transformed MNIST images. It is important to note that the traditional machine learning concepts of training and testing sets do not apply here, as the solver does not require label knowledge of the MNIST images – all the information it processes comes solely from the sensor measurements (and initial temperature).

Finally, to evaluate the reconstruction performance, we compare the results obtained from the moving sensor setup with those from an alternative configuration involving 16 and 64 static sensors, respectively. The comparison in the video https://youtu.be/Z9h16V2FRss allows us to assess the impact of sensor mobility on the accuracy of the conductivity reconstruction.

moving sensors

Figure 7 illustrates the forward problem using the transformed MNIST[0] image as atrutha_{\mathrm{truth}}. We obtain the temperature dynamics on a sufficiently dense time grid in t∈[0,1]t \in[0,1] and show only the representative snapshots at t=0t = 0, t=12t = \frac{1}{2}, t=1t = 1 respectively in Figure 7b. Next, according to the configuration in Figure 2, the 4 moving sensors record temperature in Figure 8a. Then, we initiate the gradient descent with a constant initial guess (top-left of Figure 8c), which serves as a conservative lower bound for the true conductivity. Figure 8d is the simulated measurement taking areconstructeda_{\mathrm{reconstructed}} as the constant initial guess, but there is a significant discrepancy when compared to Figure 8a, and their differences are displayed in Figure 8e. The gradient of the loss function, depicted in the bottom-left of Figure 8c, is used to guide the descent direction. By following this gradient over 500 epochs, we reach the final estimate, resulting in the reconstructed conductivity areconstructeda_{\mathrm{reconstructed}}, shown in the top-right of Figure 8c.

Image not extracted

Figure 8. (a) Temperature measurements from 4 moving sensors; (b) Loss and error decay; (c) Reconstruction and gradient of loss; (d) Initial simulated measurement; (e) Initial residual; (f) Final simulated measurement; (g) Final residual. Training log on MNIST[0]

The simulated measurement taking the final estimate as areconstructeda_{\mathrm{reconstructed}} is shown in Figure 8g, and the corresponding residual Figure 8g is much smaller, indicating improved accuracy in the reconstruction process.

Baseline: 16 and 64 equidistant static sensors

For a fair comparison, the setup of forward problem is exactly the same as the 4 moving sensors, and the only difference is the placement of the sensors, isolating the impact of sensor configuration on the measurement and reconstruction performance. Notice that only two curves appear in Figure 9e because all measurements overlap to them due to the symmetry in initial condition, sensor positions and the initial guess areconstructeda_{\mathrm{reconstructed}} (top-left of Figure 9d).

The reconstruction using 16 static sensors is quite poor (top-right of Figure 9d). While increasing the number of static sensors to 64 significantly improves the reconstruction accuracy (top-right of Figure 10d), the trade-off is the substantially higher number of sensors required.

Image not extracted

Figure 10. 64 static sensors

To compare three different configurations of sensor placements effectively, we focus on evaluating the reconstruction error at a given loss level, rather than tracking the training epoch. Specifically, we examine how the reconstruction error typically behaves as the loss decreases, as shown in Figure 11. The primary interval of interest is for loss values below 10−610^{-6}, as gradient descent generally minimizes the loss to this range (Figure 8b, Figure 9c and Figure 10c). Notably, the worst-case performance of the 4 moving sensors still outperforms the best-case scenario of 16 static sensors in terms of relative reconstruction error, despite recording no more amount of data. Furthermore, since the final residual is at the level of 10−310^{-3} (Figure 8g, Figure 9h and Figure 10h), measurements are noise-resilient up to the same scale of additive noise. The code for T² is available at https://github.com/yugt/inverseheat.

Figure 11. Loss error relation based on first 21 images of MNIST

Conclusion

In this work, we present a novel and efficient method for reconstructing thermal conductivity in both 1D and 2D domains using moving sensors. Our approach leverages the mobility of sensors to maximize spatial coverage during the most sensitive phase of temperature measurement, allowing for accurate conductivity reconstruction with fewer sensors. The use of automatic differentiation significantly reduces computational costs, enabling scalable and robust conductivity recovery. Extensive simulations on both 1D and 2D domains, including evaluations on transformed MNIST

16 sensor positions

(a) 16 sensor positions

(b) Sensor measurements

(c) Loss and error decay

Reconstruction and gradient of loss

(d) Reconstruction and gradient of loss

Initial simulated measurement

(e) Initial simulated measurement

Initial residual

(f) Initial residual

Final simulated measurement

(g) Final simulated measurement

(h) Final residual

Figure 9. 16 static sensors

images, demonstrate the method’s effectiveness. Notably, even with only four moving sensors, the reconstruction accuracy outperforms configurations with 16 or even 64 static sensors, indicating the advantage of sensor mobility in sparse data collection scenarios. The proposed approach, combining reduced computational burden and improved reconstruction quality, offers a practical and versatile solution for inverse problems in heat equations, applicable to a wide range of real-world thermal conductivity mapping tasks. Future work may extend these techniques to domains with complex boundary conditions or multi-scale conductivity variations, further enhancing the applicability of this method.

Acknowledgements

The author sincerely thanks the referees for their useful comments. The author also extends gratitude to Agnid Banerjee for proving lemma 2.2, and to Kookjin Lee for providing access to a private computing node on the ASU Sol Supercomputer [7], without which this work will take thousands of CPU hours serially. The author acknowledge the usage of ChatGPT to fix grammar and general style and to perform proof-reading.

Animation

https://youtu.be/Z9h16V2FRss shows the video of training logs on multiple instances of MNIST image as ground truth conductivity. All the 4 input instance are tested on the three methods: 4 moving sensors, 16 equidistant static sensors and 64 equidistant static sensors. The snapshots are also shown in appendix C.

Code Snippets

The code for the 1D circle is primarily pedagogical, serving as a foundation and inspiration for applying similar methods to 2D torus. In this implementation, self.ode_func corresponds to the ODE

u˙(t)=A(a)u(t)+f(t),(15)\dot{u}(t) = A(a)u(t) + f(t), \tag*{(15)}

while the external heat source self.heat_source is defined as in

f(t,x)=sin⁡(πt),(t,x)∈[0,∞)×T.(16)f(t,x) = \sin(\pi t), \qquad(t,x) \in[0,\infty) \times\mathbb{T}. \tag*{(16)}

self.forward_PDE first fills the matrix A(a)A(a) based on the input a(⋅)a(\cdot) on T\mathbb{T}, and its output represents u(t,x)u(t,x) on [0,1]×T[0,1] \times\mathbb{T}, with each temperature measurement u(tm,xj)u(t_m,x_j) having automatic differentiation enabled.

In contrast, the code for the 2D torus is thoroughly tested, well-maintained, and highly recommended as the go-to reference. It is hosted on https://github.com/yugt/inverse-heat.

Listing 1: InverseHeat1D

import torch
from torchdiffeq import odeint

torch.backends.cudnn.deterministic = True
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

class InverseHeat1D:
    fs2t = lambda self, fm: self.fbasis[:, :fm.shape[0]] @ fm

    def __init__(self, J=100, M=None, fmodes=19) -> None:
if M is None:
    M = J
self.t = torch.linspace(0, 1, M+1).to(device)
self.x = torch.linspace(0, 1, J+1)[:-1].to(device)
self.f = torch.sin(torch.pi * self.t).unsqueeze(-1) @\
    torch.ones_like(self.x).unsqueeze(0)
self.u0 = torch.sin(2 * torch.pi * self.x)
self.fbasis = torch.stack([
    torch.cos(- * 2*torch.pi * self.x) if - % 2==0
    else torch.sin(- * 2*torch.pi * self.x)
    for - in range(fmodes) ]).T.to(device)
self.fbasis /= torch.sqrt(torch.diag(self.fbasis.T @ self.fbasis))

def heat_source(self, t, x):
    return torch.sin(torch.pi * t) * torch.ones_like(x)

def ode_func(self, t, x, A=None):
    return A @ x + self.heat_source(t, x)  # $u'(t) = A(a)u(t) + f(t)$

def forward_PDE(self, a=None):
    if a is None:
        a = self.a
    A = torch.diag(a[:-1], diagonal=1) +\
        torch.diag(a[:-1], diagonal=-1)
    A[0, -1] = a[-1]
    A[-1, 0] = a[-1]
    A -= torch.diag(a + torch.roll(a, shifts=1))
    # Matrix A(a) is prepared
    return odeint(lambda t, u: self.ode_func(t, u,
                             A=A * self.x.shape[0]**2),
                  self.u0, self.t)

def formulate(a:torch.Tensor, name:str):
    solver = InverseHeat1D(J=100, fmodes=19)
    solver.a = a
    m = torch.zeros((solver.t.shape[0], solver.x.shape[0]),
                    dtype=torch.bool)
    for _ in range(1, m.shape[0]):
        m[-, -1] = True
    solver.sensor = m
    solver.observe = solver.forward_PDE()[solver.sensor]
    solver.name = name
    return solver

def trainer(solver, epoch=500):
    log = {'loss':[], 'error':[], 'fbasis':[]}
    gradescendent = lambda params: torch.optim.SGD([params], lr=50,
                              momentum=0.1, dampening=0.1, weight_decay=0)
    ## steepest descent
    afm = -41*torch.ones(1).to(device)
    afm.requires_grad = True
    l2 = torch.nn.MSELoss(reduction='mean').to(device)
    opt = gradescent(afm)
    for _ in range(epoch):
        opt.zero_grad()
        ahat = torch.exp(solver.fs2t(afm))
        predict = solver.forward_PDE(ahat)[solver.sensor]
        loss = l2(solver.observe, predict)
loss.backward()
opt.step()
error = ((ahat - solver.a)**2).sum() / solver.a.shape[0]
print(afm.grad.detach().cpu(),
      afm.grad.detach().cpu(),
      loss.detach(), error.detach())
log['loss'].append(torch.log10(loss).item())
log['error'].append(torch.log10(error).item())
log['fbasis'].append(afm.shape[0])
if (afm.grad**2).sum() < 1e−7 and
    afm.shape[0] < solver.fbasis.shape[1]:
    temp = torch.zeros(afm.shape[0] + 1).to(device)
    temp[:-1] = afm.detach()
    afm = temp.requires_grad_()
opt = gradescent(afm)

Extra figures for 2D

The following figures mentions “4m” in their titles, which is an abbreviation of “4 moving”.

Figure 12. (a) Temperature dynamics and measurements from 4m sensors. (b) Initial guess: constant conductivity. (c) Reconstruction after 500 GD steps. Traning log for MNIST[04].

Figure 13. (a) Temperature dynamics and measurements from 16 sensors

(b) Initial guess: constant conductivity

(c) Reconstruction after 500 GD steps

Figure 13: Traning log for MNIST[04]

Figure 14. (a) Temperature dynamics and measurements from 64 sensors; (b) Initial guess: constant conductivity; (c) Reconstruction after 500 GD steps. Figure 14: Traning log for MNIST [04]

Figure 15. Traning log for MNIST[12]

Figure 16. (a) Temperature dynamics and measurements from 16 sensors

(b) Initial guess: constant conductivity

(c) Reconstruction after 500 GD steps

Figure 16: Traning log for MNIST[12]

Figure 17. (a) Temperature dynamics and measurements from 64 sensors; (b) Initial guess: constant conductivity; (c) Reconstruction after 500 GD steps. Traning log for MNIST[12]

Training log plots for MNIST

Figure 18. (a) Temperature dynamics and measurements from 4m sensors. (b) Initial guess: constant conductivity. (c) Reconstruction after 500 GD steps. Figure 18: Traning log for MNIST [13]

Figure 19. (a) Temperature dynamics and measurements from 16 sensors; (b) Initial guess: constant conductivity; (c) Reconstruction after 500 GD steps. Traning log for MNIST[13]

Figure 20. (a) Temperature dynamics and measurements from 64 sensors. (b) Initial guess: constant conductivity. (c) Reconstruction after 500 GD steps. Figure 20: Traning log for MNIST[13]

References

  1. [1]Alexandros Beskos, Mark Girolami, Shiwei Lan, Patrick E. Farrell, and Andrew M. Stuart. Geometric mcmc for infinite-dimensional inverse problems. Journal of Computational Physics, 335:327–351, 2017. ISSN 0021-9991. doi: https://doi.org/10.1016/j.jcp.2016.12.041. URL https://www.sciencedirect.com/science/article/pii/S0021999116307033.
  2. [3]Franck Boyer and Pierre Fabrie. Mathematical Tools for the Study of the Incompressible Navier-Stokes Equations and Related Models, volume 183. Springer New York, 2013. ISBN 978-1-4614-5974-3. doi: 10.1007/978-1-4614-5975-0.DOI
  3. [4]K. Cao and D. Lesnic. Reconstruction of the space-dependent perfusion coefficient from final time or time-average temperature measurements. Journal of Computational and Applied Mathematics, 337:150–165, 2018. ISSN 0377-0427. doi: https://doi.org/10.1016/j.cam.2018.01.010. URL https://www.sciencedirect.com/science/article/pii/S0377042718300311.
  4. [5]Ricky T. Q. Chen. torchdiffeq, 2018. URL https://github.com/rtqichen/torchdiffeq.
  5. [6]Shukai Du and Samuel N Stechmann. Inverse radiative transfer with goal-oriented hp-adaptive mesh refinement: adaptive-mesh inversion. Inverse Problems, 39(11):115002, sep 2023. doi: 10.1088/1361-6420/acf785.DOI
  6. [7]Douglas M. Jennew ein, Johnathan Lee, Chris Kurtz, Will Dizon, Ian Shaeffer, Alan Chapman, Alejandro Chiquete, Josh Burks, Amber Carlson, Natalie Manson, Arhat Kobwala, Thirugnana Jagadeesan, Praful Barghav, Torey Battelle, Rebecca Belshe, Debra McCaffrey, Marisa Brazil, Chaitanya Inumella, Kirby Kuznia, Jade Buzinski, Sean Dudley, Dhruvil Shah, Gil Speyer, and Jason Yalim. The Sol Supercomputer at Arizona State University. In Practice and Experience in Advanced Research Computing, PEARC ’23, pages 296–301, New York, NY, USA, Jul 2023. Association for Computing Machinery. ISBN 9781450399852. doi: 10.1145/3569951.3597573.
  7. [8]Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
  8. [9]Ming-Ding Li, Xiao-Quan Shen, Xin Chen, Jia-Ming Gan, Fang Wang, Jian Li, Xiao-Liang Wang, and Qun-Dong Shen. Thermal management of chips by a device prototype using synergistic effects of 3-d heat-conductive network and electrol caloric refrigeration. Nature Communications, 13(1):5849, Oct 2022. ISSN 2041-1723. doi: 10.1038/s41467-022-33596-z. URL https://doi.org/10.1038/s41467-022-33596-z.
  9. [10]Y Nagasaka and A Nagashima. Absolute measurement of the thermal conductivity of electrically conducting liquids by the transient hot-wire method. Journal of Physics E: Scientific Instruments, 14(12):1435, dec 1981. doi: 10.1088/0022-3735/14/12/020. URL https://dx.doi.org/10.1088/0022-3735/14/12/020.
  10. [11]Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zach DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In Proceedings of the 31st International Conference on Neural Information Processing Systems (NeurIPS 2017), pages 2572–2583, Long Beach, California, USA, 2017. Curran Associates, Inc. URL https://api.semanticscholar.org/CorpusID:40027675.
  11. [12]Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kӧpf, Edward Yang, Zach DeVito, Martin Raison, Al ykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. ArXiv, abs/1912.01703, 2019. URL https://api.semanticscholar.org/CorpusID:202786778.arxiv.org/abs/1912.01703
  12. [14]David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning representations by back-propagating errors. Nature, 323(6088):533–536, 1986. ISSN 1476-4687. doi: 10.1038/323533a0. URL https://doi.org/10.1038/323533a0.
  13. [15]Arthur Sard. The measure of the critical values of differentiable maps. Bulletin of the American Mathematical Society, 48(12):883 – 890, 1942.DOI
  14. [16]A. M. Stuart. Inverse problems: A bayesian perspective. Acta Numerica, 19:451–559, 2010. doi: 10.1017/S0962492910000061.DOI
  15. [17]TorchVision maintainers and contributors. Torchvision: Pytorch’s computer vision library. https://github.com/pytorch/vision, 2016.
  16. [18]G.K. Wallace. The jpeg still picture compression standard. IEEE Transactions on Consumer Electronics, 38(1):xviii–xxxiv, 1992. doi: 10.1109/30.125072.DOI

Paper details

vlnc:W_znyjjvbs
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.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents