betaFeedback

WorksMathematics

Prediction Uncertainty Controlled Computer Model Calibration

Abstract

Computer-model calibration is usually undertaken in service of prediction, yet neither Bayesian Kennedy-O'Hagan (KOH) calibration nor amortized simulation-based inference is built to control the predictive error of the calibrated simulator. When parameters are weakly identified or the data-model discrepancy is misspecified, the simulator-output bands induced by a nominal posterior level can be far from nominal. We propose a two-layer framework that targets predictive miscoverage directly. First, a conditional flow-matching model learns an amortized posterior for the calibration parameters, trained on simulator outputs perturbed with Gaussian-process discrepancies of varying structure. Second, conformal risk control arranges posterior draws for the target observation into nested candidate sets and selects the smallest one whose fixed-rank pointwise bands attain a prescribed time-averaged miscoverage on a synthetic calibration sample. Under exchangeability of calibration and target observations, the bands control expected miscoverage in finite samples. In a synthetic benchmark, the proposed bands are roughly 20% the width of KOH-MCMC bands with miscoverage near the 10% target. In calibrating an age-structured SEIR model to the 2010 Malawi measles outbreak, they are again sharper and near-nominal, and the weeks they fail to cover coincide with the asymmetric epidemic decline, exposing structural inadequacy of the mechanistic model.

Introduction

Computer models—deterministic simulators of complex physical, biological, and engineering systems—are indispensable scientific instruments in fields ranging from epidemiology and climate science to hydrology and astrophysics. Their practical utility depends on calibration: learning unknown simulator inputs from observations so that the calibrated simulator can support prediction, intervention design, and scientific inference [19, 4, 16, 7, 32]. These downstream tasks require more than accurate parameter point estimates; they require predictive uncertainty that is neither misleadingly narrow nor unnecessarily diffuse. For example, Lim et al. [23] developed a more accurate method for estimating epidemic initial conditions but identified wide credible intervals and below-nominal empirical coverage as limitations requiring further work. Motivated by this concern, we focus on predictive uncertainty for simulator outputs: intervals should accurately identify scientifically relevant trajectories.

An increasingly attractive approach to calibration is to learn the inverse map from simulator outputs to input parameters. This inverse-model approach has been developed for high-dimensional dependent outputs using deep neural networks [5] and can be viewed within the broader framework of simulation-based inference (SBI), in which simulations rather than a tractable likelihood provide the information needed for inference [10]. Modern neural SBI methods train a conditional density estimator on simulated parameter–output pairs and thereby amortize the cost of inference: after a single training stage, the fitted model can rapidly return an approximate posterior for each new observation [30, 29, 12], which makes SBI compelling for repeated, sequential, or time-sensitive calibration problems.

Amortization, however, does not by itself make posterior predictive uncertainty reliable. Under weak parameter identifiability, data–model discrepancy, or approximation error in the learned inverse map, some calibration parameters can be recovered imprecisely, and propagating the resulting parameter distribution through the simulator can produce predictive intervals that are excessively wide, too narrow, or otherwise poorly aligned with their nominal level [15]. Standard SBI procedures approximate parameter distributions; they do not control the predictive error of their simulator pushforwards. The conventional Kennedy–O’Hagan (KOH) Bayesian formulation [19] shares this predictive limitation while introducing a computational one: its posterior is commonly explored by Markov chain Monte Carlo (MCMC), which can be expensive and must generally be rerun for every new observation [5, 10]. Moreover, nominal posterior probability levels and the repeated-sample coverage of simulator-output bands are different quantities, and parameter–discrepancy nonidentifiability or misspecification can further compromise inference [7, 38]. Figure 1 previews how KOH–MCMC and a naïve SBI construction can both yield conservative simulator-output bands for the same measles outbreak. We consequently take amortized SBI as our methodological starting point and use MCMC-based KOH calibration as a benchmark.

KOH–MCMC posterior simulator-output band

(a) KOH–MCMC posterior simulator-output band.

Flow-matching SBI with fixed posterior-density retention and no CRC

(b) Flow-matching SBI with fixed posterior-density retention and no CRC.

Figure 1. Predictive bands for the 2010 Malawi measles outbreak. Panel (a) shows the nominal 90% pointwise KOH–MCMC band: empirical miscoverage is 3.45% on the evaluated weeks and average pointwise width is 3,526.89 cases. Panel (b) shows a flow-matching SBI band obtained by retaining the highest-density 90% of posterior parameter draws: empirical miscoverage is 0% and average width is 6,078.54 cases. Blue dashed curves are medians of simulator outputs, green dotted curves are band limits, and red solid curves are observed weekly cases; orange shading marks missed observations when present. Miscoverage is evaluated only at weeks with at least 300 observed cases.

Our central contribution is to introduce prediction-accuracy-based conformal risk control (CRC) into neural SBI. We combine CRC [3] with flow-matching posterior estimation (FMPE) [11]—a scalable neural posterior estimator based on continuous normalizing flows [24] with competitive performance on established SBI benchmarks [11, 26]. Rather than accepting a fixed posterior probability region, CRC selects the calibrated parameter set by directly evaluating a predictive miscoverage loss after parameter draws are propagated through the simulator, identifying the region of the parameter space most relevant for accurate prediction. Under exchangeability between the CRC calibration examples and a future target, the resulting simulator-output intervals control expected miscoverage in finite samples. This prediction-driven correction is related in spirit to flow-based conformal forecasting [22], but our target is the simulator pushforward of an amortized parameter posterior in a computer-model calibration problem. This prediction-driven calibration is practical only because the posterior is amortized: CRC requires an approximate posterior for each of hundreds of synthetic calibration observations conditioned on the target, which amounts to a few hundred ODE integrations with a trained flow but would demand as many separate MCMC runs under the standard KOH workflow, repeated for every new observation. Amortization is therefore not merely a computational convenience but the condition that makes prediction-accuracy-based risk control feasible.

We also address the discrepancy sensitivity of the inverse map. Following the learning-with-noise strategy of Bhatnagar et al. [5], we train FMPE on simulator outputs contaminated by draws from a flexible heterogeneous Gaussian-process discrepancy family and generate the separate CRC calibration sample under the corresponding discrepancy mechanism. The framework therefore combines rapid amortized parameter inference with a second, prediction-focused layer that corrects the size of the simulator-output set, and it can be applied to SBI methods other than FMPE.

The remainder of the paper is organized as follows. Section 2 reviews KOH and simulation-based calibration, conformal risk control, and flow matching. Section 3 develops the proposed framework. Section 4 reports a simulation study using the synthetic model of Bhatnagar et al. [5], Section 5 applies the framework to the Malawi measles outbreak, and Section 6 concludes.

Background

Kennedy–O’Hagan (KOH) Calibration and Simulation-Based Inference

The standard KOH calibration framework, proposed by Kennedy and O’Hagan [19] and developed in many subsequent works (see, e.g., [4, 16, 39, 32, 33]), is based on the observation model

Z=Y(θ)+δ,(1)Z = Y(\theta) + \delta, \tag*{(1)}

where θ∈Θ⊂Rd\theta\in\Theta\subset\mathbb{R}^{d} denotes the calibration parameter vector; Y(θ)∈RTY(\theta) \in\mathbb{R}^{T} is the deterministic simulator output at TT discrete time points or spatial locations; Z∈RTZ \in\mathbb{R}^{T} denotes the corresponding observational data; and δ∈RT\delta\in\mathbb{R}^{T} is a data–model discrepancy term capturing systematic mismatch between the simulator and reality, including observational errors. Given a prior p(θ)p(\theta) and the likelihood p(Z∣θ)p(Z \mid\theta) induced by (1), KOH calibration targets the posterior p(θ∣Z)∝p(Z∣θ)p(θ)p(\theta\mid Z) \propto p(Z \mid\theta)p(\theta), which quantifies parametric uncertainty; propagating posterior draws through Y(θ)Y(\theta) yields a posterior distribution for the simulator output. This forward formulation is statistically delicate: θ\theta and the discrepancy can be nonidentifiable without careful prior specification or experimental design, and ignoring or misspecifying the discrepancy can bias inference on the physical parameters [7, 39, 32]. It is also computationally demanding, since a sufficiently long MCMC chain must be run for the observation at hand and repeated when a new observation arrives.

A separate issue concerns the predictive meaning of the resulting bands. A nominal posterior probability level describes uncertainty under the assumed model; it does not, by itself, guarantee that a band covers observed outcomes at the same rate under repeated sampling [38], and conventional KOH calibration contains no step that tunes simulator-output bands to a prespecified predictive-miscoverage loss. Figure 1a illustrates this with the KOH–MCMC baseline for the Malawi measles outbreak analyzed in § 5. Its nominal 90% pointwise band has empirical miscoverage 0.0345 among weeks with at least 300 reported cases (coverage of approximately 96.55%). The broad band, particularly around the epidemic peak, illustrates the cost in sharpness when posterior uncertainty is propagated without an explicit predictive-risk calibration step.

The inverse-model approach reverses the computational direction and learns calibration parameters directly from simulator outputs; Bhatnagar et al. [5], for example, train a deep neural network on simulated parameter–output pairs for time-series calibration. More broadly, inference from an implicit simulator is described as simulation-based inference [10]. In neural posterior estimation, one simulates pairs (θ,x)(\theta, x) from the joint model and fits a conditional density qφ(θ∣x)q_{\varphi}(\theta\mid x) to approximate p(θ∣x)p(\theta\mid x) [29, 12]; the training cost is paid once, after which qφ(θ∣Z)q_{\varphi}(\theta\mid Z) can be sampled for any new observation at negligible additional cost. SBI nevertheless changes the computational mechanism, not the predictive objective. If a component of θ\theta is weakly identified, the learned conditional distribution is inaccurate, or the observed discrepancy differs from those represented during training, the approximate posterior may be displaced or overly dispersed; passing all of its draws through Y(θ)Y(\theta) transfers that error to the output scale. A nominal credible region from KOH–MCMC and a nominal probability region from SBI therefore share a crucial limitation: neither is constructed to guarantee that the corresponding simulator-output interval attains its nominal predictive coverage (see, e.g., Hermans et al., [15]; Syring and Martin [38]).

Figure 1b makes the SBI limitation concrete using approximate posterior samples from a trained flow-matching model for the same Malawi outbreak. We retain the 90% of generated parameter draws with the highest estimated conditional posterior density, pass them through the SEIR simulator, and form pointwise bands from the forward trajectories. The band misses none of the evaluated weeks, but its average pointwise width is 6,078.54 cases, indicating overcoverage (see Section S4 of the Supplementary Material). This distinction motivates the two-layer structure of our framework: flow-matching SBI supplies rapid approximate parameter draws, while CRC chooses among the induced simulator-output sets using prediction error on a separate synthetic calibration sample, targeting the quantity that matters downstream—predictive miscoverage—with explicit finite-sample risk control under exchangeability.

Conformal Risk Control

Conformal risk control (CRC) extends split conformal prediction from ensuring marginal coverage to controlling a user-specified expected loss, retaining finite-sample validity under exchangeability (Angelopoulos et al., [3]; Vovk et al. [41]; Romano et al. [35]; Angelopoulos and Bates [2]). Let Dcal:={(Xi,Yi)}i=1n\mathcal{D}_{\mathrm{cal}} := \{(X_i,Y_i)\}_{i=1}^{n} denote a separate calibration dataset and (Xn+1,Yn+1)(X_{n+1},Y_{n+1}) a new test point, where XiX_i represents the input features and YiY_i the response. Throughout, we use the term calibration in two senses: computer model calibration (learning simulator parameters θ\theta) and CRC calibration (using a separate sample Dcal\mathcal{D}_{\mathrm{cal}} to tune statistical risk).

Let Λ⊆R\Lambda\subseteq\mathbb{R} be a nonempty, totally ordered set of admissible threshold values whose largest element λmax⁡=sup⁡Λ\lambda_{\max} = \sup\Lambda exists, and let {C^λ}λ∈Λ\{\widehat{C}_{\lambda}\}_{\lambda\in\Lambda} be a family of prediction sets in which larger λ\lambda yields more conservative sets. For each observation ii, define a bounded loss

Li(λ):=L(C^λ(Xi),Yi),Li(λ)∈(−∞,B],L_i(\lambda) := L\left(\widehat{C}_{\lambda}(X_i),Y_i\right), \qquad L_i(\lambda) \in(-\infty,B],

where BB is a known upper bound; Li(λ)L_i(\lambda) measures the discrepancy between C^λ(Xi)\widehat{C}_{\lambda}(X_i) and YiY_i (e.g., empirical miscoverage). The goal is a data-dependent threshold λ^\widehat{\lambda} such that E[Ln+1(λ^)]≤α\mathbb{E}[L_{n+1}(\widehat{\lambda})] \leq\alpha for a target level α∈(0,1)\alpha\in(0,1). Assuming each Li(λ)L_i(\lambda) is non-increasing and right-continuous in λ\lambda with Li(λmax⁡)≤αL_i(\lambda_{\max}) \leq\alpha almost surely, let R^n(λ):=n−1∑i=1nLi(λ)\widehat{R}_n(\lambda) := n^{-1}\sum_{i=1}^{n} L_i(\lambda) be the empirical calibration risk and

select

λ^=inf⁡{λ∈Λ:nn+1R^n(λ)+Bn+1≤α}=inf⁡{λ∈Λ:R^n(λ)≤αadj},(2)\hat{\lambda}=\inf\left\{\lambda\in\Lambda:\frac{n}{n+1}\widehat{R}_n(\lambda)+\frac{B}{n+1}\leq\alpha\right\}=\inf\left\{\lambda\in\Lambda:\widehat{R}_n(\lambda)\leq\alpha_{\mathrm{adj}}\right\}, \tag*{(2)}

where αadj=(n+1)α/n−B/n\alpha_{\mathrm{adj}}=(n+1)\alpha/n-B/n is the finite-sample adjusted target level, with λ^=λmax⁡\hat{\lambda}=\lambda_{\max} if the set is empty. Angelopoulos et al. [3] show that if Li(λ)L_i(\lambda), i=1,…,n+1i=1,\ldots,n+1, are exchangeable, then E[Ln+1(λ^)]≤α\mathbb{E}[L_{n+1}(\hat{\lambda})]\leq\alpha. In our setting, CRC thus provides a rigorous mechanism for constructing prediction intervals with controlled expected miscoverage when nominal credible levels cannot be trusted.

Flow Matching

We approximate the target distribution of the calibration parameters using flow matching, which learns a time-dependent vector field transporting a simple reference distribution to a complex target ([24]). We use the Gaussian conditional probability path

pt(x∣z)=N(x;tz,(1−(1−σmin⁡)t)2Id),t∈[0,1],p_t(x\mid z)=\mathcal{N}\left(x;tz,\left(1-(1-\sigma_{\min})t\right)^2I_d\right),\qquad t\in[0,1],

where x∈Rdx\in\mathbb{R}^d is the state variable, z∼pdataz\sim p_{\mathrm{data}} is a target sample, and σmin⁡∈(0,1)\sigma_{\min}\in(0,1) is a small variance floor. The trajectory starts from standard Gaussian noise at t=0t=0 and concentrates around zz as t→1t\to1, so that marginally p0(x)=N(0,Id)p_0(x)=\mathcal{N}(0,I_d) and p1(x)≈pdatap_1(x)\approx p_{\mathrm{data}}. A neural network vtϕ(x)v_t^\phi(x) with weights ϕ\phi is trained with the conditional flow matching (CFM) objective

LCFM(ϕ)=Et∼π, z∼pdata, x∼pt(⋅∣z)[∥vtϕ(x)−vtflow(x∣z)∥2],vtflow(x∣z)=z−(1−σmin⁡)x1−(1−σmin⁡)t,\mathcal{L}_{\mathrm{CFM}}(\phi)=\mathbb{E}_{t\sim\pi,\ z\sim p_{\mathrm{data}},\ x\sim p_t(\cdot\mid z)}\left[\left\|v_t^\phi(x)-v_t^{\mathrm{flow}}(x\mid z)\right\|^2\right],\qquad v_t^{\mathrm{flow}}(x\mid z)=\frac{z-(1-\sigma_{\min})x}{1-(1-\sigma_{\min})t},

where π(t)\pi(t) is a sampling distribution over [0,1][0,1] and vtflowv_t^{\mathrm{flow}} is the closed-form conditional vector field generating the Gaussian path. This objective is a tractable, simulation-free surrogate for learning the target transport ([24, 25]); it is minimized by stochastic gradient descent on minibatch Monte Carlo estimates without solving an ODE during training. After training, approximate samples are generated by integrating the learned ordinary differential equation (ODE)

X0∼N(0,Id),dXt=vtϕ(Xt) dt,t∈[0,1],X_0 \sim\mathcal{N}(0,I_d), \qquad\mathrm{d}X_t = v_t^\phi(X_t)\,\mathrm{d}t, \qquad t \in[0,1],

and taking the terminal state X1X_1.

Because our goal is to sample from the conditional posterior given observed covariates yy (e.g., simulator outputs), we condition the vector field on yy and employ classifier-free guidance (CFG) at sampling time,

v~t(x∣y)=(1−wt)vtϕ(x∣∅)+wtvtϕ(x∣y),wt>1,\tilde{v}_t(x \mid y) = (1-w_t)v_t^\phi(x \mid\varnothing) + w_t v_t^\phi(x \mid y), \qquad w_t > 1,

where ∅\varnothing denotes a learnable unconditional null-embedding [42] (Ho and Salimans 2022): on condition-dropped training examples the raw condition y=[A,Z]∈Ra+Ty = [A,Z] \in\mathbb{R}^{a+T}, with a=dim⁡(A)a = \dim(A), is replaced by a trainable vector e∅∈Ra+Te_\varnothing\in\mathbb{R}^{a+T} that is optimized jointly with the network, so the conditional and unconditional fields share all other parameters. A general formulation of flow matching, including marginal and conditional paths, objective equivalence, and guided flows, is provided in Section S1.2 of the Supplementary Material.

Prediction Uncertainty Controlled Calibration

This section develops our framework for constructing simulation-based calibration sets with finite-sample control over expected predictive miscoverage. Throughout, the observed sequence, simulator output, and discrepancy have common dimension TT: for θ∈Θ\theta\in\Theta, Y(θ)=(Y1(θ),…,YT(θ))⊤Y(\theta) = (Y_1(\theta), \ldots, Y_T(\theta))^\top, δ=(δ1,…,δT)⊤\delta= (\delta_1, \ldots, \delta_T)^\top, and Z=Y(θ)+δ=(Z1,…,ZT)⊤∈RTZ = Y(\theta) + \delta= (Z_1, \ldots, Z_T)^\top\in\mathbb{R}^T. When the loss is evaluated only at scientifically relevant coordinates, I(Z)⊆{1,…,T}\mathcal{I}(Z) \subseteq\{1, \ldots, T\} denotes their deterministic index set and T(Z)=∣I(Z)∣\mathcal{T}(Z) = |\mathcal{I}(Z)|; otherwise I(Z)={1,…,T}\mathcal{I}(Z) = \{1, \ldots, T\} and T(Z)=T\mathcal{T}(Z) = T. Let α∈(0,1)\alpha\in(0,1) denote the target risk level, nn the size of the CRC calibration dataset, and NN the number of amortized posterior draws used to construct each candidate parameter set.

To tune the conformal parameter sets, we construct a synthetic CRC calibration dataset Dn:={(A~i,Z~i,θi)}i=1nD_n := \{(\widetilde{A}_i,\widetilde{Z}_i,\theta_i)\}_{i=1}^n that mimics the joint distribution of the parameters and the observables: parameters θi\theta_i are drawn from the trained flow-matching posterior approximation conditioned on the target observation, and synthetic observations Z~i=Y(θi)+δ~i\widetilde{Z}_i = Y(\theta_i) + \widetilde{\delta}_i are formed with simulated discrepancies δ~i\tilde{\delta}_i drawn from a sufficiently flexible distribution, since exact knowledge of the discrepancy law is not assumed. In some problems we optionally extract auxiliary summaries A~i=A(θi)∈A\tilde{A}_i = A(\theta_i) \in\mathcal{A}, forward-model outputs assumed to be observed exactly. The construction of Dn\mathcal{D}_n is detailed in § 3.4.2. § 3.1 presents the miscoverage loss and parameter-set constructions, § 3.3 the conditional flow-matching architecture, and § 3.4 the discrepancy-simulation strategy.

Miscoverage Risk

For the ii-th calibration sample, define Ii:=I(Z~i)\mathcal{I}_i := \mathcal{I}(\tilde{Z}_i) and Ti:=∣Ii∣T_i := |\mathcal{I}_i|. Let Λ⊆R\Lambda\subseteq\mathbb{R} be a nonempty, totally ordered set of admissible thresholds with largest element λmax⁡=sup⁡Λ\lambda_{\max} = \sup\Lambda. Denote the candidate parameter set given an auxiliary summary AA and observed data ZZ by Cλ(A,Z)\mathcal{C}_{\lambda}(A,Z), λ∈Λ\lambda\in\Lambda, and require the family to be nested: λ≤λ′\lambda\leq\lambda' implies Cλ(A,Z)⊆Cλ′(A,Z)\mathcal{C}_{\lambda}(A,Z) \subseteq\mathcal{C}_{\lambda'}(A,Z). Let mi,λ:=∣Cλ(A~i,Z~i)∣m_{i,\lambda} := |\mathcal{C}_{\lambda}(\tilde{A}_i,\tilde{Z}_i)| and let Niref:=mi,λmax⁡≥1N_i^{\mathrm{ref}} := m_{i,\lambda_{\max}} \geq1 be the reference set size. Define the fixed lower and upper tail ranks by

ki−:=⌈Nirefρ2⌉,qi+:=max⁡{1,⌊Niref(1−ρ2)⌋},ki+:=Niref−qi++1,k_i^- := \left\lceil N_i^{\mathrm{ref}}\frac{\rho}{2}\right\rceil,\qquad q_i^+ := \max\left\{1,\left\lfloor N_i^{\mathrm{ref}}\left(1-\frac{\rho}{2}\right)\right\rfloor\right\},\qquad k_i^+ := N_i^{\mathrm{ref}}-q_i^+ + 1,

where ρ∈(0,1)\rho\in(0,1) is a user-specified interval-tail hyperparameter; ki−k_i^- is the rank from the lower tail and ki+k_i^+ the rank from the upper tail. These ranks are computed once from NirefN_i^{\mathrm{ref}} and reused for every candidate λ\lambda.

For any finite parameter set Θ\Theta and coordinate uu, let Yu,(r)(Θ)Y_{u,(r)}(\Theta), 1≤r≤∣Θ∣1 \leq r \leq|\Theta|, denote the rr-th order statistic of {Yu(θ):θ∈Θ}\{Y_u(\theta):\theta\in\Theta\}. For a candidate set with enough retained draws, i.e., mi,λ≥ki−+ki+−1m_{i,\lambda} \geq k_i^- + k_i^+ - 1, define the fixed-rank prediction interval at coordinate uu by [bˉiu(λ),biu+(λ)][\bar{b}_{iu}(\lambda),b^+_{iu}(\lambda)], where bˉiu(λ):=Yu,(ki−)(Cλ(A~i,Z~i))\bar{b}_{iu}(\lambda) := Y_{u,(k_i^-)}(\mathcal{C}_{\lambda}(\tilde{A}_i,\tilde{Z}_i)), ri,λ+:=mi,λ−ki++1r^+_{i,\lambda} := m_{i,\lambda}-k_i^+ + 1, and biu+(λ):=Yu,(ri,λ+)(Cλ(A~i,Z~i))b^+_{iu}(\lambda) := Y_{u,(r^+_{i,\lambda})}(\mathcal{C}_{\lambda}(\tilde{A}_i,\tilde{Z}_i)). The fixed-rank empirical miscoverage loss is then defined as follows. If Ti=0T_i = 0, set Li(λ)=0L_i(\lambda)=0. If Ti>0T_i > 0 but mi,λ<ki−+ki+−1m_{i,\lambda} < k_i^- + k_i^+ - 1, set Li(λ)=1L_i(\lambda)=1. Otherwise,

Li(λ):=1Ti∑u∈IiI{Z~iu∉[bˉiu(λ),biu+(λ)]}.(3)L_i(\lambda) := \frac{1}{T_i}\sum_{u\in\mathcal{I}_i}\mathbb{I}\left\{\tilde{Z}_{iu}\notin\left[\bar{b}_{iu}(\lambda),b^+_{iu}(\lambda)\right]\right\}. \tag*{(3)}

The fixed-rank construction guarantees monotonicity. Because the candidate sets are nested in λ\lambda, adding draws can only decrease the ki−k_i^--th smallest and increase the ki+k_i^+-th largest simulated value, so biu(λ)b^{\mathrm{iu}}(\lambda) is non-increasing and biu(λ)b^{\mathrm{iu}}(\lambda) non-decreasing in λ\lambda and the intervals expand monotonically with λ\lambda. Consequently the threshold selected by CRC in equation (2) favors the narrowest bands that satisfy the empirical constraint.

We introduce two formulations for the candidate sets Cλ(A,Z)C_{\lambda}(A,Z). For each conditioning pair (A,Z)(A,Z), consider the conditional ODE of § 2.3, dXt=vtϕ(Xt∣A,Z)dt\mathrm{d}X_t=v_t^{\phi}(X_t\mid A,Z)\mathrm{d}t, X0=x0X_0=x_0, and let ψ0,tϕ(x0∣A,Z):=Xt\psi_{0,t}^{\phi}(x_0\mid A,Z):=X_t denote the induced flow map, with set-valued pushforward ψ0,tϕ(U∣A,Z):={ψ0,tϕ(u∣A,Z):u∈U}\psi_{0,t}^{\phi}(U\mid A,Z):=\{\psi_{0,t}^{\phi}(u\mid A,Z):u\in U\} for finite U⊂RdU\subset\mathbb{R}^d. In both formulations, increasing λ\lambda expands the set, driving down the fixed-rank miscoverage loss.

(i) Radius-based set. This approach truncates the base distribution prior to the flow transformation. Let SN={ξj}j=1NS_N=\{\xi_j\}_{j=1}^{N} be an ensemble of NN base samples ξj∼iidN(0,Id)\xi_j\mathrel{\overset{\mathrm{iid}}{\sim}}\mathcal{N}(0,I_d) and Bλ(0)={x∈Rd:∥x∥2≤λ}B_{\lambda}(0)=\{x\in\mathbb{R}^d:\|x\|_2\le\lambda\} the L2L_2-ball of radius λ\lambda. We define

Cλ(A,Z):=ψ0,1ϕ(Bλ(0)∩SN∣A,Z),C_{\lambda}(A,Z):=\psi_{0,1}^{\phi}\left(B_{\lambda}(0)\cap S_N\mid A,Z\right),

with admissible threshold domain Λrad=[0,max⁡1≤j≤N∥ξj∥2]\Lambda_{\mathrm{rad}}=[0,\max_{1\le j\le N}\|\xi_j\|_2], whose largest element retains the complete base ensemble.

(ii) Posterior-based set. Alternatively, we threshold the conditional posterior density. Let ΘNϕ(A,Z):=ψ0,1ϕ(SN∣A,Z)\Theta_N^{\phi}(A,Z):=\psi_{0,1}^{\phi}(S_N\mid A,Z) be the generated parameter ensemble and, for any finite set Θ\Theta, let Qρ{f(θ):θ∈Θ}Q_{\rho}\{f(\theta):\theta\in\Theta\} denote the empirical ρ\rho-quantile of ff over Θ\Theta. The posterior-based set retains the highest-density draws up to λ∈[0,1]\lambda\in[0,1]:

Cλ(A,Z):={θ∈ΘNϕ(A,Z):log⁡p1ϕ(θ∣A,Z)≥Q1−λ{log⁡p1ϕ(ϑ∣A,Z):ϑ∈ΘNϕ(A,Z)}},C_{\lambda}(A,Z):=\left\{\theta\in\Theta_N^{\phi}(A,Z):\log p_1^{\phi}(\theta\mid A,Z)\ge Q_{1-\lambda}\left\{\log p_1^{\phi}(\vartheta\mid A,Z):\vartheta\in\Theta_N^{\phi}(A,Z)\right\}\right\},

with Λpos=[0,1]\Lambda_{\mathrm{pos}}=[0,1]; in practice, Cλ(A,Z)C_{\lambda}(A,Z) consists of the top ⌈λN⌉\lceil\lambda N\rceil generated draws ranked by conditional log-density. The conditional log-density is computed along the probability flow ODE via the continuous normalizing-flow (CNF) identity [9, 24]

ddtlog⁡ptϕ(Xt∣A,Z)=−div⁡vtϕ(Xt∣A,Z),(4)\frac{d}{dt}\log p_t^{\phi}(X_t\mid A,Z)=-\operatorname{div}v_t^{\phi}(X_t\mid A,Z), \tag*{(4)}

where XtX_t traces the trajectory from X0∼p0ϕ(θ∣Aobs,Zobs)=N(0,Id)X_0\sim p_0^{\phi}(\theta\mid A_{\mathrm{obs}},Z_{\mathrm{obs}})=\mathcal{N}(0,I_d) to X1=θ∼p1ϕ(θ∣Aobs,Zobs)X_1=\theta\sim p_1^{\phi}(\theta\mid A_{\mathrm{obs}},Z_{\mathrm{obs}}).

Selection of the trimming parameter.

After fitting the flow, the tail hyperparameter ρ\rho is selected separately for each construction j∈{rad,pos}j \in\{\mathrm{rad},\mathrm{pos}\} using simulated tuning data generated by the mechanism of § 3.4.2, independently of the final CRC calibration sample: a tuning CRC sample of size ntune=100n_{\mathrm{tune}} = 100 and a validation sample of size nv=500n_v = 500. For each ρ\rho on the grid G={0.01,0.02,…,0.20}\mathcal{G} = \{0.01, 0.02, \ldots, 0.20\}, we compute a tuning threshold λ^jtune(ρ)\hat{\lambda}^{\mathrm{tune}}_j(\rho) from equation (2) with ntunen_{\mathrm{tune}} in place of nn, and evaluate it on the validation sample using two criteria: the validation miscoverage and the mean pointwise interval length,

R^j,val(ρ)=1nv∑v=1nvLjvval(λ^jtune(ρ);ρ),\widehat{R}_{j,\mathrm{val}}(\rho) = \frac{1}{n_v}\sum_{v=1}^{n_v} L^{\mathrm{val}}_{jv}\left(\hat{\lambda}^{\mathrm{tune}}_j(\rho);\rho\right),
Wj(ρ)=1nvT∑v=1nv∑u=1T[bjvu+(λ^jtune(ρ);ρ)−bjvu−(λ^jtune(ρ);ρ)].(5)W_j(\rho) = \frac{1}{n_vT}\sum_{v=1}^{n_v}\sum_{u=1}^{T}\left[b^+_{jvu}\left(\hat{\lambda}^{\mathrm{tune}}_j(\rho);\rho\right)-b^-_{jvu}\left(\hat{\lambda}^{\mathrm{tune}}_j(\rho);\rho\right)\right]. \tag*{(5)}

where LjvvalL^{\mathrm{val}}_{jv} is the loss in equation (3) for validation observation vv and bjvu±b^{\pm}_{jvu} are the corresponding interval endpoints at coordinate uu; we set Wj(ρ)=+∞W_j(\rho) = +\infty if any validation interval lacks enough retained draws to be defined. The first criterion ensures that the tuned procedure attains the target risk on fresh data, and the second measures sharpness. We select the admissible value with the sharpest intervals,

ρ^j∈arg⁡min⁡ρ∈GjadmWj(ρ),Gjadm={ρ∈G:R^j,val(ρ)≤α, Wj(ρ)<∞},(6)\hat{\rho}_j \in\arg\min_{\rho\in\mathcal{G}^{\mathrm{adm}}_j} W_j(\rho), \qquad\mathcal{G}^{\mathrm{adm}}_j = \left\{\rho\in\mathcal{G}: \widehat{R}_{j,\mathrm{val}}(\rho)\leq\alpha,\ W_j(\rho)<\infty\right\}, \tag*{(6)}

and the rule yields no selection if Gjadm\mathcal{G}^{\mathrm{adm}}_j is empty. We then freeze ρj=ρ^j\rho_j = \hat{\rho}_j, generate a fresh final CRC calibration sample, and recompute the final threshold λ^j\hat{\lambda}_j. Conditional on the fitted model and tuning information, the CRC bound retains its form provided the assumptions of § 2.2 hold for the final calibration and target losses [3].

Deep Approximate Distribution with Flow Matching

We parameterize the conditional vector field vtϕ(x∣A,Z)v^\phi_t(x \mid A, Z) by a deep neural network with inputs (x,A,Z,t)(x, A, Z, t), where x∈Rdx \in\mathbb{R}^d is the latent state at time t∈[0,1]t \in[0,1] [11, 24, 25]. The conditioning information (t,A,Z)(t, A, Z) enters through a context vector ct=[ht,hZ,A]c_t = [h_t, h_Z, A].

The time embedding ht∈R256h_t \in\mathbb{R}^{256} is obtained by passing random Fourier features of tt [34] through a two-layer SiLU network. The sequence representation hZh_Z is built by dividing Z∈RTZ \in\mathbb{R}^{T} into overlapping patches of width 7 and stride 2, linearly embedding each patch into R128\mathbb{R}^{128}, prepending a learned start-of-sequence token, adding positional encodings, and passing the result through a four-layer Transformer encoder [40]. The encoder output is aggregated by learned-query multihead-attention pooling with q=4q = 4 queries and 4 heads [21], projected back to R128\mathbb{R}^{128}, and augmented with two geometric features, log⁡{1+max⁡uZu}\log\{1 + \max_u Z_u\} and the normalized location of the maximum, giving hZ∈R130h_Z \in\mathbb{R}^{130}.

Given xx and ctc_t, the backbone consists of Nblk=4N_{\mathrm{blk}} = 4 fixed-width residual blocks of hidden dimension 256 with Adaptive Layer Normalization (AdaLN) and Swish-Gated Linear Units (SwiGLU) [36, 31]: in each block, the context vector produces a scale and shift applied to the layer-normalized state, which is then passed through a gated linear unit with dropout and added residually. The final normalized state is concatenated with ctc_t and passed through dd coordinate-specific output heads to produce vtϕ(x∣A,Z)∈Rdv_t^{\phi}(x \mid A, Z) \in\mathbb{R}^{d}. The context projections, residual output projections, and output heads are initialized to zero, so each block initially acts as the identity and the initial vector field is zero, which improves numerical stability early in training. The complete layer-by-layer specification is provided in Section S1.1 of the Supplementary Material.

Handling Data-Model Discrepancy

Data-model discrepancy—a mismatch between the model output Y(θ)Y(\theta) and the observed data ZZ even at the best input setting—is a central challenge in computer model calibration, and ignoring it can bias inference on θ\theta [5, 7, 39]. To make the flow-matching model robust to discrepancy, we train it on model outputs contaminated with synthetic discrepancy terms sampled from Gaussian processes with varying hyperparameters. Exposure to a diverse array of discrepancies encourages the learned inverse map to identify features that are informative for recovering the physical parameters while accounting for discrepancy.

Training Data Generation

We first draw a pool of NparamN_{\mathrm{param}} parameter configurations θk\theta_k, k∈{1,…,Nparam}k \in\{1,\ldots,N_{\mathrm{param}}\}, via Latin Hypercube Sampling (LHS) over the prior bounds [27] and evaluate the simulator to obtain the clean trajectories Y(θk)Y(\theta_k) and auxiliary outputs A~k=A(θk)\tilde{A}_k = A(\theta_k). Separately, we generate a pool of NnoiseN_{\mathrm{noise}} raw discrepancy realizations δ~j\tilde{\delta}_j, j∈{1,…,Nnoise}j \in\{1,\ldots,N_{\mathrm{noise}}\}, from a zero-mean Gaussian process (GP) [5] whose marginal variance scales with the temporal variability of the clean output. Writing σY2(θ)=1T∑u=1T(Yu(θ)−Yˉ(θ))2\sigma_Y^2(\theta) = \frac{1}{T}\sum_{u=1}^{T}(Y_u(\theta)-\bar{Y}(\theta))^2 with Yˉ(θ)=1T∑u=1TYu(θ)\bar{Y}(\theta) = \frac{1}{T}\sum_{u=1}^{T}Y_u(\theta), the covariance between indices u1,u2∈{1,…,T}u_1,u_2 \in\{1,\ldots,T\} is

K(u1,u2;θ)=σY2(θ)[ζI(u1=u2)+κexp⁡(−∣u1−u2∣τ)].(7)K(u_1,u_2;\theta) = \sigma_Y^2(\theta)\left[\zeta\mathbb{I}(u_1=u_2)+\kappa\exp\left(-\frac{|u_1-u_2|}{\tau}\right)\right]. \tag*{(7)}

The prefactor σY2(θ)\sigma_Y^2(\theta) keeps the discrepancy magnitude commensurate with the amplitude of each trajectory, so that the nugget ζ\zeta, partial sill κ\kappa, and range τ\tau control the noise-to-signal ratio, the relative strength of the temporally correlated component, and the correlation length, respectively. These hyperparameters are sampled via LHS from user-specified bounds for each realization. The training set pairs every parameter configuration kk with every noise realization jj,

Z~k,j,u=Yu(θk)+δ~j,u,(8)\tilde{Z}_{k,j,u} = Y_u(\theta_k) + \tilde{\delta}_{j,u}, \tag*{(8)}

with contaminated states clipped to their theoretical domain (e.g., nonnegative counts), yielding Dtrain={(A~k,Z~k,j,θk)}\mathcal{D}_{\mathrm{train}} = \{(\tilde{A}_k,\tilde{Z}_{k,j},\theta_k)\} of size Ntrain=Nparam×NnoiseN_{\mathrm{train}} = N_{\mathrm{param}} \times N_{\mathrm{noise}}.

Calibration Data Generation for Conformal Risk Control

After training, the nn CRC calibration samples in Dn\mathcal{D}_n are generated independently, conditional on the fitted flow and the observed conditioning pair. We draw θi∼p1ϕ(θ∣Aobs,Zobs)\theta_i \sim p_1^{\phi}(\theta\mid A_{\mathrm{obs}}, Z_{\mathrm{obs}}) via ODE integration, each requiring only a forward pass through the trained flow rather than a new posterior computation, clamp the draws to their physical bounds, and pass each through the simulator to obtain Y(θi)Y(\theta_i) and A~i\tilde{A}_i. Rather than reusing the training noise pool, we independently sample new GP hyperparameters and generate a unique raw GP draw δ~i\tilde{\delta}_i for each θi\theta_i, and then apply a nonnegative, problem-specific state-dependent envelope Mu{Y(θi)}M_u\{Y(\theta_i)\} to reproduce the heteroscedastic scale pattern of the target observation:

Z~i,u=Yu(θi)+Mu{Y(θi)}δ~i,u,u=1,…,T.(9)\widetilde{Z}_{i,u} = Y_u(\theta_i) + M_u\{Y(\theta_i)\}\widetilde{\delta}_{i,u}, \qquad u=1,\ldots,T. \tag*{(9)}

The scale at coordinate uu may depend on the entire clean trajectory; both studies below use

Mu{Y(θ)}={Yu(θ)−min⁡1≤s≤TYs(θ)max⁡1≤s≤TYs(θ)−min⁡1≤s≤TYs(θ)}1/2,(10)M_u\{Y(\theta)\} = \left\{\frac{Y_u(\theta)-\min_{1\le s\le T}Y_s(\theta)}{\max_{1\le s\le T}Y_s(\theta)-\min_{1\le s\le T}Y_s(\theta)}\right\}^{1/2}, \tag*{(10)}

so that discrepancies are attenuated in low-signal regions. After modulation, the sequence is clipped to its physical bounds. The synthetic calibration samples are therefore independent draws from

Pcal(A,Z,θ)=pδ,M{Z∣Y(θ)}I{A=A(θ)}p(θ),(11)P_{\mathrm{cal}}(A,Z,\theta)=p_{\delta,M}\{Z\mid Y(\theta)\}\mathbb{I}\{A=A(\theta)\}p(\theta), \tag*{(11)}

where p(θ)=p1ϕ(θ∣Aobs,Zobs)p(\theta)=p_1^\phi(\theta\mid A_{\mathrm{obs}},Z_{\mathrm{obs}}) and pδ,Mp_{\delta,M} is the conditional law induced jointly by the raw GP, the envelope MM, and clipping.

The two discrepancy mechanisms serve different purposes. Unmodulated discrepancies in Dtrain\mathcal{D}_{\mathrm{train}} are a data-augmentation device that makes the learned inverse map robust to severe discrepancies across the full trajectory, whereas the envelope in equation (9) makes the synthetic calibration data reproduce the state-dependent variability anticipated in the observed data, encouraging exchangeability of the target observation with the calibration data. The finite-sample CRC conclusion rests on the working assumption that the target observation (Aobs,Zobs)(A_{\mathrm{obs}},Z_{\mathrm{obs}}) and the nn synthetic observations {(A~i,Z~i)}i=1n\{(\widetilde{A}_i,\widetilde{Z}_i)\}_{i=1}^n are jointly exchangeable under this envelope-modulated mechanism. Since the loss Li(λ)L_i(\lambda) depends only on the observations and not on θi\theta_i, the CRC result of § 2.2 applies under this assumption [3].

Simulation Study

We evaluate the proposed framework using the synthetic computer model and observations of Bhatnagar et al. [5]. Unlike the application in § 5, the synthetic output Y(θ)Y(\theta) is available in closed form without a numerical solver, providing an inexpensive testbed. Throughout this section, the target miscoverage level is α=0.1\alpha=0.1.

Synthetic Model Outputs and Observational Data

The deterministic forward model produces smooth unimodal curves governed by θ=(θ1,θ2,θ3)T∈[0,1]3\boldsymbol{\theta} = (\theta_{1}, \theta_{2}, \theta_{3})^{\mathsf{T}} \in[0,1]^{3}. For T=53T = 53 time points s(u)=4(u−1)T−1−2s(u) = \frac{4(u-1)}{T-1} - 2 equally spaced on [−2,2][-2,2], the output at index uu is

Yu(θ)=0.3+θ1+0.32π(θ3+0.1)exp⁡[−(s(u)−θ2+0.5)2θ3+0.1].Y_{u}(\boldsymbol{\theta}) = 0.3 + \frac{\theta_{1} + 0.3}{\sqrt{2\pi(\theta_{3} + 0.1)}} \exp\left[-\frac{(s(u) - \theta_{2} + 0.5)^{2}}{\theta_{3} + 0.1}\right].

where θ1\theta_{1} controls the amplitude, θ2\theta_{2} the peak location, and θ3\theta_{3} the dispersion. Representative outputs and the synthetic observation are shown in Figure 2.

Synthetic observation and representative model outputs from the simulation-study forward model

Figure 2. Synthetic observation and representative model outputs from the simulation-study forward model.

All datasets share the GP discrepancy family of § 3.4, with the kernel in equation (7) and hyperparameters sampled via LHS from range proportion τprop∈[0.05,0.20]\tau_{\mathrm{prop}} \in[0.05, 0.20] (so that τ=τprop⋅T\tau= \tau_{\mathrm{prop}} \cdot T), partial sill κ∈[0.052,0.252]\kappa\in[0.05^{2}, 0.25^{2}], and nugget ζ∈[0.0052,0.0252]\zeta\in[0.005^{2}, 0.025^{2}].

Training data. We draw Ntrain=5,000N_{\mathrm{train}} = 5{,}000 parameter configurations via LHS over [0,1]3[0,1]^{3} and Nnoise=30N_{\mathrm{noise}} = 30 GP hyperparameter vectors, and pair every configuration with every noise realization via equation (8), yielding 150,000 contaminated trajectories. As discussed in § 3.4, the training discrepancies are left unscaled by any state-dependent envelope so that the network sees severe discrepancies at all signal levels.

Synthetic observation data. We fix a ground-truth parameter θ∗\boldsymbol{\theta}^{*} and draw a raw GP discrepancy ϵobs\epsilon_{\mathrm{obs}} with covariance given by equation (7) at θ∗\boldsymbol{\theta}^{*}, modulated by the envelope of equation (10):

Zobs,u=Yu(θ∗)+Mu{Y(θ∗)}ϵobs,u,u=1,…,T.(12)Z_{\mathrm{obs},u} = Y_{u}(\boldsymbol{\theta}^{*}) + M_{u}\{Y(\boldsymbol{\theta}^{*})\}\epsilon_{\mathrm{obs},u}, \qquad u = 1,\ldots,T. \tag*{(12)}

Calibration data. A separate CRC calibration dataset Dn\mathcal{D}_n of size n=100n = 100 is built following § 3.4.2: parameters are sampled through the trained flow map ψ0,1ϕ(⋅∣Aobs,Zobs)\psi^{\phi}_{0,1}(\cdot\mid A_{\mathrm{obs}}, Z_{\mathrm{obs}}) and clamped to [0,1]3[0,1]^3, and each draw receives a unique envelope-modulated GP discrepancy as in equation (9). The synthetic target and the CRC calibration observations thus deliberately share the same discrepancy mechanism (GP family, hyperparameter law, and envelope), isolating performance under the nominal specification; matching the discrepancy mechanisms does not by itself establish joint exchangeability, which remains a working assumption (§ 3.4.2). The thresholds λ^rad\hat{\lambda}_{\mathrm{rad}} and λ^pos\hat{\lambda}_{\mathrm{pos}} are calibrated once on Dn\mathcal{D}_n and held fixed thereafter.

Implementation and Evaluation Protocol

We use the conditional vector-field architecture of § 3.3 with σmin⁡=0.1\sigma_{\min} = 0.1. Posterior samples are generated by integrating the conditional ODE with the Euler scheme over 100 equally spaced steps on [0,1][0,1], using classifier-free guidance scale w=1.2w = 1.2 with an interval-based scheduler (Ho and Salimans, 2022, Zheng et al., [42], Kynkää n niemi et al., [20]), and the generated parameters are mapped back to the original scale and clamped to [0,1]3[0,1]^3. For both set constructions we draw N=100N = 100 latent base samples ξj∼iidN(0,Id)\xi_j \mathrel{\overset{\mathrm{iid}}{\sim}} \mathcal{N}(0,I_d) and transport them through the flow; the same latent pool is shared across all nn calibration samples so that differences among parameter sets are driven by the conditioning observations rather than base-space sampling noise.

The miscoverage loss of equation (3) uses the independently selected values ρrad=ρpos=0.03\rho_{\mathrm{rad}} = \rho_{\mathrm{pos}} = 0.03 (§ 3.2) and the relevant-coordinate set I(Z)={u:Zu≥η}\mathcal{I}(Z) = \{u : Z_u \ge\eta\} with η=0.301\eta= 0.301. Since the clean simulator has baseline 0.3 and minimum attainable peak 0.3+0.3/2π(1.1)≈0.4140.3+0.3/\sqrt{2\pi}(1.1) \approx0.414 over [0,1]3[0,1]^3, η\eta lies only 0.001 (about 0.88% of the minimum dynamic range) above the baseline and excludes only essentially flat background coordinates. For the radius-based sets, the ODE trajectories of all NN latent draws conditioned on each calibration observation are precomputed together with the base-space norms ∥ξj∥2\lVert\xi_j\rVert_2, reducing the threshold search to a filtering operation; for the posterior-based sets, the conditional log-densities computed via equation (4) are cached alongside the generated parameters. In both cases λ^\hat{\lambda} is obtained by applying Brent’s root-finding algorithm (Brent, [6]) to Lˉ(λ)−αadj\bar{L}(\lambda) - \alpha_{\mathrm{adj}}, where Lˉ(λ)=n−1∑i=1nLi(λ)\bar{L}(\lambda) = n^{-1}\sum_{i=1}^{n} L_i(\lambda) and αadj=(n+1)α/n−1/n\alpha_{\mathrm{adj}} = (n+1)\alpha/n - 1/n; the search is carried out over λ∈(0,max⁡1≤j≤N∥ξj∥2]\lambda\in(0,\max_{1\le j\le N}\lVert\xi_j\rVert_2] for the radius-based sets and over λ∈(0.001,0.999)\lambda\in(0.001,0.999) for the posterior-based sets, with tolerance 10−410^{-4}.

We use repeated held-out simulation to assess how each procedure behaves for repeated measurements of the same system. Holding θ∗\theta^{*} fixed, we conduct Nexp⁡=200N_{\exp}=200 experiments, regenerating an independent envelope-modulated GP discrepancy with newly sampled hyperparameters according to equation (12) in each. The CRC thresholds are estimated once from DnD_n and frozen before the held-out experiments, mimicking deployment to future observations. For each held-out observation we construct parameter sets of maximum size Nref=100N^{\mathrm{ref}}=100 and the associated predictive bands, and record (i) empirical miscoverage, (ii) average pointwise interval length (sharpness), and (iii) RMSE between the held-out observation and the band median (center accuracy). For comparison, a KOH–MCMC baseline is rerun separately for every held-out observation and its pointwise (1−α)(1-\alpha) posterior band is constructed, reproducing the observation-specific, non-amortized Bayesian workflow (MCMC specifications and convergence diagnostics are given in § 13.1 of the Supplementary Material).

Results

Figure 3 displays the predictive bands produced by the radius-based, posterior-based, and MCMC methods for a single representative synthetic observation. Here, we intentionally show a case where the radius-based and posterior-based methods lead to almost the same result. The blue dashed curve is the median of the simulator outputs over the parameter set, the green dotted curves trace the pointwise 90% prediction interval, the red solid curve is the synthetic observation, and orange shading marks time points where the observation falls outside the interval (and exceeds η\eta). On this draw, the MCMC band envelops the observation entirely at the cost of roughly 7×7\times longer pointwise intervals and roughly twice the RMSE between the predictive median and the observation, whereas the two CRC bands enclose the observation much more tightly.

Prediction bands for the radius-based construction

(a) Radius-based construction: miscoverage 0.0943, RMSE 0.0119, average pointwise interval length 0.0337.

Prediction bands for the posterior-based construction

(b) Posterior-based construction: miscoverage 0.0943, RMSE 0.0118, average pointwise interval length 0.0339.

Prediction bands for the MCMC baseline

(c) MCMC baseline: miscoverage 0.0000, RMSE 0.0232, average pointwise interval length 0.2058.

Figure 3. Prediction bands for the same representative synthetic observation under the two CRC constructions and the KOH–MCMC baseline, with target miscoverage α=0.1\alpha= 0.1. Line and shading conventions are described in the text.

Figure 4 summarizes the three metrics across the Nexp⁡=200N_{\exp}=200 experiments. The radius-based set, calibrated at λ^rad=1.3078\hat{\lambda}_{\mathrm{rad}}=1.3078, attains a mean miscoverage of 0.1222, close to the target, with some dispersion (75% quantile 0.1989). The posterior-based set, calibrated to retain the top λ^pos=46.47%\hat{\lambda}_{\mathrm{pos}}=46.47\% of draws by conditional log-density, is more conservative with mean miscoverage 0.0721 (75% quantile 0.1132). The MCMC baseline attains zero miscoverage in all 200 experiments.

Three boxplots showing (a) Miscoverage, (b) RMSE, and (c) Average interval length

Figure 4. (a) Miscoverage (b) RMSE (c) Average interval length. Distributions of evaluation metrics across 200 independent experiments: empirical miscoverage, RMSE between the predictive median and the synthetic observation, and average pointwise interval length. The horizontal dashed line in panel (a) marks the target level α=0.1\alpha= 0.1.

ments, reflecting bands wide enough to envelop almost every realization regardless of how the discrepancy is distributed. The RMSE and interval-length panels make the cost of this overcoverage explicit: the CRC methods achieve mean RMSE 0.0089 (radius) and 0.0085 (posterior), roughly a third of the MCMC mean of 0.0222, and mean pointwise interval lengths of 0.0363 and 0.0426, about 20% of the MCMC mean of 0.2073—a pattern consistent with parameter–discrepancy nonidentifiability widening posterior uncertainty [7].

These results illustrate the central trade-off targeted by the framework. The MCMC bands reflect posterior uncertainty under an assumed Gaussian observation model but are not calibrated to a prespecified predictive-risk target [38]; they are over-wide on average and cannot adapt to the realized discrepancy. The CRC methods calibrate the interval width directly against the empirical miscoverage on synthetic data generated under the same flexible GP discrepancy family [3], producing intervals that are tight, adaptive, and near the nominal target. An ablation that bypasses the CRC step and instead retains a fixed 90% of the generated draws under either selection rule (Section S4 of the Supplementary Material) indicates that this behavior is driven by the CRC calibration step rather than by the set constructions themselves: for the representative observation of Figure 3, both fixed-retention rules attain zero miscoverage but produce average interval lengths of 0.0597, about 1.8 times those of the corresponding CRC bands, indicating overcoverage.

Application to a Measles SEIR Model

We now apply the framework to a real-data calibration problem: inferring the parameters of an age-structured Susceptible–Exposed–Infected–Recovered (SEIR) model from the 2010 measles outbreak in Malawi [28]. This application is more challenging than the simulation study in three respects. First, the forward model is a system of 84 coupled ODEs whose evaluation is far more expensive than the closed-form curve. Second, the posterior is non-identifiable along certain directions in parameter space, complicating both inference and discrepancy modeling. Third, the data are real, so the true data-generating mechanism is unknown and the discrepancy assumption underlying the exchangeability argument can only be assessed indirectly. Our goal is to demonstrate finite-sample miscoverage control under these realistic conditions while remaining computationally favorable relative to the Bayesian baseline.

Measles Outbreak Data and Age-Structured SEIR Model

The data, originally analyzed by Minetti et al. [28] and shown in Figure 5, consist of reported weekly measles case counts ZobsZ_{\mathrm{obs}} at T=53T=53 epochs spanning 52 weeks (364 days) and the empirical proportions of cases across three coarse age groups (0.5–5, 5–15, and 15+ years), Aobs=(0.42,0.30,0.28)∈Δ2⊂R3A_{\mathrm{obs}}=(0.42, 0.30, 0.28)\in\Delta^2\subset\mathbb{R}^3, where Δ2\Delta^2 is the 2-simplex. The time series ZobsZ_{\mathrm{obs}} is treated as the noisy observation subject to discrepancy, while AobsA_{\mathrm{obs}} plays the role of the auxiliary summary AA, assumed to be observed exactly.

Observed measles data and representative SEIR model outputs

Figure 5. Observed measles data and representative SEIR model outputs.

We adopt an age-structured SEIR formulation [37] in which the host population is partitioned into K=21K = 21 age classes (a six-month band for 0.5–1 year, one-year bands up to 19–20 years, and a single class for 20+ years). For each class ii, let SiS_i, EiE_i, IiI_i, RiR_i denote the number of susceptible, exposed, infectious, and recovered individuals. The dynamics are

dSidt=−λi(t)Si,dEidt=λi(t)Si−σEi,dIidt=σEi−γIi,dRidt=γIi,(13)\begin{aligned} \frac{dS_i}{dt} &= -\lambda_i(t) S_i, & \frac{dE_i}{dt} &= \lambda_i(t) S_i - \sigma E_i, \\ \frac{dI_i}{dt} &= \sigma E_i - \gamma I_i, & \frac{dR_i}{dt} &= \gamma I_i, \tag*{(13)} \end{aligned}

with latent and infectious periods both fixed at seven days, σ=γ=1/7\sigma= \gamma= 1/7 [37]. The age-specific force of infection is density dependent,

λi(t)=β∑j=1KWijIj(t),(14)\lambda_i(t) = \beta\sum_{j=1}^{K} W_{ij} I_j(t), \tag*{(14)}

where β>0\beta> 0 is a per-pair transmission rate and W=(Wij)W = (W_{ij}) is the Who-Acquires-Infection-From-Whom (WAIFW) matrix. Following Anderson and May [1], Shea et al. [37], WijW_{ij} depends only on ∣i−j∣\lvert i-j\rvert, decreasing from 1 on the diagonal to 2/3 for ∣i−j∣≥4\lvert i-j\rvert\ge4 (the full specification is given in § 12 of the Supplementary Material), encoding contact rates that are highest within an age class and decline with age separation.

At t=0t = 0, the susceptible age distribution is a discretized Gamma density: Si(0)S_i(0) is proportional to the Gamma(ash,art)(a_{\mathrm{sh}}, a_{\mathrm{rt}}) density at the class midpoint, normalized so that ∑i=1KSi(0)=N0\sum_{i=1}^{K} S_i(0) = N_0, the pre-outbreak susceptible population size. A small fraction fE∈(0,1)f_E \in(0,1) of each class is moved into the exposed compartment to seed the epidemic, Ei(0)=max⁡{fE⋅Si(0),1}E_i(0) = \max\{f_E \cdot S_i(0), 1\}, with Si(0)S_i(0) adjusted accordingly and Ii(0)=Ri(0)=0I_i(0) = R_i(0) = 0. Because the elapsed time between epidemic onset and the first reporting week is unknown, we introduce a warm-up duration t0>0t_0 > 0: the system is integrated from t=0t = 0 to t=t0t = t_0, and the simulator is then evaluated at t=t0+7(u−1)t = t_0 + 7(u - 1) days for u=1,…,Tu = 1, \ldots, T. The resulting weekly case-count sequence is the simulator output Y(θ)∈RTY(\theta) \in\mathbb{R}^{T}, and the time-summed age-group proportions of the infectious compartment over the observation window,

ag=∑i∈Gg∫t0t0+364Ii(t) dt∑i=1K∫t0t0+364Ii(t) dt,g∈{1,2,3},a_g = \frac{\sum_{i \in\mathcal{G}_g} \int_{t_0}^{t_0+364} I_i(t)\,\mathrm{d}t}{\sum_{i=1}^{K} \int_{t_0}^{t_0+364} I_i(t)\,\mathrm{d}t}, \qquad g \in\{1,2,3\},

with Gg\mathcal{G}_g the index sets of the three age groups, form the auxiliary summary $A(\theta) = (a_1,a_2,a_3) \in

Δ2\Delta^2. The calibration target is the six-dimensional vector θ=(β,N0,fE,ash,art,t0)T\boldsymbol{\theta} = (\beta, N_0, f_E, a_{\mathrm{sh}}, a_{\mathrm{rt}}, t_0)^\mathsf{T}, sampled from a Latin hypercube design with the bounds in Table 1.

ParameterDefinitionPrior support
β\betaPer-pair transmission rate (day−1^{-1})[0.5×10−6, 2.5×10−6][0.5 \times10^{-6},\ 2.5 \times10^{-6}]
N0N_0Susceptible population size[1.4×105, 3×105][1.4 \times10^5,\ 3 \times10^5]
fEf_EInitial exposed fraction[10−4, 10−3][10^{-4},\ 10^{-3}]
asha_{\mathrm{sh}}Shape of age distribution[0.8, 1.4][0.8,\ 1.4]
arta_{\mathrm{rt}}Rate of age distribution[0.09, 0.39][0.09,\ 0.39]
t0t_0Warm-up duration (days)[0.1, 10][0.1,\ 10]

Table 1. Latin hypercube bounds for the SEIR calibration parameters.

Discrepancy, Synthetic Data, and Implementation

We assume that the discrepancy between Y(θ)Y(\boldsymbol{\theta}) and the observed weekly cases is generated by the heterogeneous GP family of § 3.4, with hyperparameters sampled via LHS from bounds calibrated to the scale of the measles data: τprop∈[0.1, 0.3]\tau_{\mathrm{prop}} \in[0.1,\,0.3], κ∈[0.102, 0.402]\kappa\in[0.10^2,\,0.40^2], and ζ∈[0.012, 0.042]\zeta\in[0.01^2,\,0.04^2]; A(θ)A(\boldsymbol{\theta}) is not subject to discrepancy. The training set is constructed exactly as in § 3.4.1, with Ntrain=5,000N_{\mathrm{train}} = 5{,}000 LHS parameter configurations from Table 1, Nnoise=30N_{\mathrm{noise}} = 30 GP hyperparameter vectors, no state-dependent envelope, and 150,000 contaminated training sequences. The CRC calibration dataset Dn\mathcal{D}_n of size n=100n = 100 follows § 3.4.2: parameters are drawn through the trained flow map conditioned on (Aobs,Zobs)(A_{\mathrm{obs}}, Z_{\mathrm{obs}}), clamped to the support in Table 1, passed through the SEIR model, and contaminated with a unique GP discrepancy modulated by the envelope of equation (10), giving independent draws from equation (11). Under the working assumption that the real-world discrepancy is itself a realization from this heteroscedastic GP family, Dn\mathcal{D}_n and the target observation are exchangeable.

The flow-matching architecture, sampling ODE, parameter-set constructions, and evaluation metrics are identical to those of § 4.2. The signal-relevance threshold is η=300\eta= 300 weekly cases, the level below which apparent miscoverage carries little epidemiological consequence, and the tuning procedure of § 3.2 selects ρrad=0.03\rho_{\mathrm{rad}} = 0.03 and ρpos=0.05\rho_{\mathrm{pos}} = 0.05. The thresholds λ^rad\widehat{\lambda}_{\mathrm{rad}} and λ^pos\widehat{\lambda}_{\mathrm{pos}} are calibrated once on Dn\mathcal{D}_n and held fixed. As a baseline, KOH calibration is run by MCMC for the real observation to sample from the posterior predictive distribution, from which pointwise (1−α)(1-\alpha) credible intervals are constructed; MCMC specifications and convergence diagnostics are given in Section S3.2 of the Supplementary Material.

MethodMiscoverageRMSEAvg. interval length
Radius-based0.1379940.772540.26
Posterior-based0.1379908.403157.48
MCMC baseline0.0345947.963526.89
Target (α=0.1\alpha= 0.1)0.1000——

Table 2. Empirical miscoverage, RMSE between the median predicted weekly cases and the observed measles outbreak, and average pointwise interval length for the radius-based parameter set, posterior-based parameter set, and MCMC baseline. The target miscoverage level is α=0.1\alpha= 0.1.

Results

Table 2 reports empirical miscoverage, RMSE between the median predicted trajectory and the observed weekly cases, and average pointwise interval length for the two proposed constructions and the MCMC baseline. Figure 6 displays the proposed bands overlaid on the Malawi data, with miscovered weeks highlighted in orange; the KOH–MCMC band was shown in Figure 1a.

Radius-based construction

(a) Radius-based construction.

Posterior-based construction

(b) Posterior-based construction.

Figure 6. CRC prediction intervals for the Malawi measles outbreak. The blue dashed line is the median of the generated SEIR trajectories, the green dotted lines are the pointwise prediction bounds, and the red solid line is the observed series. Orange bands mark weeks at which the observed value falls outside the prediction interval and exceeds the signal-relevance threshold η=300\eta= 300.

Both proposed parameter sets achieve an empirical miscoverage of 0.1379, moderately above the target α=0.1\alpha= 0.1 and consistent with the CRC guarantee E[Ln+1(λ^)]≤α\mathbb{E}[L_{n+1}(\hat{\lambda})] \leq\alpha, with average interval lengths of 2540.26 and 3157.48 cases. The MCMC baseline attains a miscoverage of 0.0345, less than half the nominal level, with the widest band (3526.89 cases); this pattern is consistent with the sensitivity of posterior uncertainty to the assumed discrepancy specification [7, 38]. The CRC procedure instead calibrates λ^\hat{\lambda} directly against the empirical miscoverage of synthetic data generated under the heterogeneous GP family, sidestepping the need for any single parametric noise model to be exactly correct. The RMSE values of the three methods are comparable. The ablation study in Section S4 of the Supplementary Material isolates the role of the CRC step on these data: retaining a fixed 90% of the generated draws without CRC, under either selection rule, attains zero miscoverage on the evaluated weeks but produces bands considerably wider than both the CRC and KOH–MCMC bands, indicating overcoverage (cf. Figure 1b).

A further feature of the framework is that the miscovered weeks, shown as orange bands in Figure 6, localize where the SEIR model struggles most. In both CRC panels the orange bands cluster around the peak and in the late tail, regions corresponding to the asymmetric falloff of the real outbreak—a pattern that the approximately symmetric, bell-shaped SEIR trajectories cannot reproduce regardless of parameter choice. The CRC framework therefore supplies not only a coverage guarantee but also a model-diagnostic signal: when the calibrated interval fails, the model is structurally inadequate. The MCMC procedure offers no analogous diagnostic because its credible intervals do not target coverage; its overcoverage reflects the same structural inadequacy but provides no statistical mechanism for surfacing it. In applications where the

validity of the predictive interval is itself of scientific interest, this combination of finite-sample guarantees and built-in diagnostics supports the proposed approach.

Discussion

This paper has developed a framework for computer model calibration that combines amortized neural posterior approximation via flow matching with conformal risk control to deliver finite-sample miscoverage guarantees on the simulator pushforward of posterior samples. In both a synthetic benchmark with known ground truth and a real-data calibration of an age-structured SEIR model to the 2010 Malawi measles outbreak, the proposed parameter sets attained near-nominal coverage with substantially sharper bands, whereas a standard KOH–MCMC baseline overcovered. The contrast reflects a difference in objective rather than in implementation quality: a Bayesian analysis under a Gaussian observation model reports credible intervals as a byproduct of posterior sampling and has limited capacity to account for structural mismatch between a mechanistic model and reality, whereas our framework targets predictive miscoverage directly and narrows the parameter set only as much as the calibration data demand.

Two features make the framework well-suited to applied calibration. First, it is amortized: the flow-matching posterior is trained once and queried rapidly for any new observation, in contrast to MCMC, which must be rerun each time [10, 11]—a consequential advantage for real-time surveillance and sequential calibration where data accrue continuously. More fundamentally, amortization is what makes the CRC layer possible at all: the calibration and tuning samples together require on the order of n+ntune+nvn+n_{\mathrm{tune}}+n_{\mathrm{v}} conditional posterior approximations per target observation, a cost that is negligible for a trained flow but prohibitive for observation-specific MCMC. Second, it supplies an interpretable diagnostic: clusters of time points at which the calibrated interval fails to cover localize the regions where the mechanistic model is structurally inadequate, directing attention toward the aspects of the model most in need of refinement.

Several limitations should be acknowledged. The finite-sample guarantee relies on exchangeability between the synthetic calibration set and the real observation [3], which holds under the working assumption that the true discrepancy lies in the user-specified GP family—a condition that cannot be verified from a single observation. The quality of the amortized posterior depends on the size and coverage of the training simulation pool, and higher-dimensional parameter spaces may require larger training budgets [10, 26]. Finally, the framework controls marginal miscoverage averaged over time; pointwise or conditional miscoverage at specific time points is not guaranteed.

Several extensions suggest themselves. Although exchangeability is untestable in the strict sense, partial diagnostics comparing summary statistics of the real observation against their empirical distribution across the calibration set—for instance, a kernel two-sample test on residual sequences [13]—would help flag gross violations. Mondrian conformal prediction [41] or localized conformal variants [14] could deliver control that adapts to local difficulty, a multi-channel extension would accommodate several observation channels with distinct discrepancy structures, and moving beyond the fixed-kernel GP family to non-stationary or non-Gaussian discrepancies, or learning the discrepancy family adaptively, would improve robustness under misspecification. More broadly, the calibration literature has long recognized discrepancy between model and reality as the principal source of inferential difficulty [19, 7, 39]. The framework presented here identifies the region of parameter space most relevant for prediction by calibrating prediction intervals against a flexible family of discrepancies, and we hope it proves useful as a robust complement to classical KOH calibration for scientific problems in which mechanistic models meet imperfect data.

Acknowledgments

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Ministry of Science and ICT (MSIT) (grant number [RS-2025-00523567,RS-2026-25607210]), the Global-LAMP Program of the NRF grant funded by the Ministry of Education of the Republic of Korea (grant number [RS-2023-00301976]).

Supplementary Material

Supplementary material available online contains the full flow-matching architecture, training details, and a general formulation of flow matching (Section S1), the WAIFW matrix (§ 12), MCMC specifications and diagnostics (§ 13), and an ablation study without conformal risk control (Section S4). The source code can be downloaded from github.com/ryanb61901/Prediction-Uncertainty-Controlled-Computer-Model-Calibration.

Data Availability Statement

The data that support the findings of this study are available from the authors upon reasonable request.

Supplementary Material

Model Details

Flow-Matching Architecture

This section gives the complete specification of the network summarized in § 3.3 of the main text. To parameterize the conditional posterior of the calibration parameters, we utilize a deep neural network to approximate the target conditional vector field vtϕ(x∣A,Z)v_t^\phi(x \mid A, Z) (Dax et al., [11]; Lipman et al., [24, 25]). Let θ∈Rd\theta\in\mathbb{R}^d denote the target calibration parameter and x∈Rdx \in\mathbb{R}^d the continuous latent state at integration time t∈[0,1]t \in[0,1]. The conditional vector field is defined as a mapping

vtϕ(⋅∣⋅,⋅):Rd×A×RT×[0,1]→Rd,(x,A,Z,t)↦vtϕ(x∣A,Z)v_t^\phi(\cdot\mid\cdot, \cdot) : \mathbb{R}^d \times\mathcal{A} \times\mathbb{R}^T \times[0,1] \to\mathbb{R}^d,\quad(x,A,Z,t) \mapsto v_t^\phi(x \mid A,Z)

governed by trainable network weights ϕ\phi. In practice, we parameterize vtϕ(x∣A,Z)v_t^\phi(x \mid A,Z) as a neural network with inputs (x,A,Z,t)(x,A,Z,t).

Let a=dim⁡(A)a = \dim(A), with a=0a = 0 when no auxiliary covariates are available. We inject information on (t,A,Z)(t,A,Z) through the context vector

ct=[ht,hZ,A]∈Rrc,c_t = [h_t,h_Z,A] \in\mathbb{R}^{r_c},

where ht∈Rrth_t \in\mathbb{R}^{r_t} is the time embedding and hZh_Z is the sequence representation defined below.

Let ω∈Rrt/2\omega\in\mathbb{R}^{r_t/2} be a vector of frequencies sampled once from N(0,σω2Irt/2)\mathcal{N}(0,\sigma_\omega^2 I_{r_t/2}) and then held fixed (Rahimi and Recht, [34]). Define

r(t)=2[sin⁡(2πtω)⊤,cos⁡(2πtω)⊤]⊤∈Rrt,r(t) = \sqrt{2}[\sin(2\pi t\omega)^\top,\cos(2\pi t\omega)^\top]^\top\in\mathbb{R}^{r_t},

and pass it through the two-layer SiLU network

ht=Wt,2SiLU⁡(Wt,1r(t)+bt,1)+bt,2∈Rrt,h_t = W_{t,2}\operatorname{SiLU}(W_{t,1}r(t) + b_{t,1}) + b_{t,2} \in\mathbb{R}^{r_t},

where Wt,1∈R2rt×rtW_{t,1} \in\mathbb{R}^{2r_t \times r_t}, Wt,2∈Rrt×2rtW_{t,2} \in\mathbb{R}^{r_t \times2r_t}, bt,1∈R2rtb_{t,1} \in\mathbb{R}^{2r_t}, and bt,2∈Rrtb_{t,2} \in\mathbb{R}^{r_t}. In our implementation, rt=256r_t = 256 and σω=16\sigma_\omega= 16 [25].

To construct hZh_Z, we first divide Z∈RTZ \in\mathbb{R}^{T} into L=⌊(T−p)/s⌋+1L = \lfloor(T-p)/s \rfloor+ 1 overlapping patches Pℓ(Z)∈RpP_\ell(Z) \in\mathbb{R}^{p} of width p=7p = 7 and stride s=2s = 2. Each patch is embedded as eℓ=WpatchPℓ(Z)+bpatch∈RrZe_\ell= W_{\mathrm{patch}}P_\ell(Z) + b_{\mathrm{patch}} \in\mathbb{R}^{r_Z}, where Wpatch∈RrZ×pW_{\mathrm{patch}} \in\mathbb{R}^{r_Z \times p}, bpatch∈RrZb_{\mathrm{patch}} \in\mathbb{R}^{r_Z}, and rZ=128r_Z = 128. After a learned start-of-sequence token in RrZ\mathbb{R}^{r_Z} is prepended and positional encodings are added, a four-layer Transformer encoder maps the sequence to

H(Z)∈R(L+1)×rZ.H(Z) \in\mathbb{R}^{(L+1) \times r_Z}.

We aggregate H(Z)H(Z) using learned-query multihead-attention pooling (Vaswani et al., [40]; Lee et al., [21]). Let Qpool∈Rq×rZQ_{\mathrm{pool}} \in\mathbb{R}^{q \times r_Z} contain q=4q = 4 learned query vectors, let Ha=4H_a = 4 be the number of attention heads, and let rhatt=rZ/Ha=32r_h^{\mathrm{att}} = r_Z/H_a = 32 be the dimension of each head. For head j∈{1,…,Ha}j \in\{1,\ldots,H_a\}, define

Oj=softmax⁡((QpoolWQ,j)(H(Z)WK,j)Trhatt)H(Z)WV,j∈Rq×rhatt.O_j = \operatorname{softmax}\left(\frac{(Q_{\mathrm{pool}}W_{Q,j})(H(Z)W_{K,j})^{\mathsf{T}}}{\sqrt{r_h^{\mathrm{att}}}}\right)H(Z)W_{V,j} \in\mathbb{R}^{q \times r_h^{\mathrm{att}}}.

where WQ,j,WK,j,WV,j∈RrZ×rhattW_{Q,j}, W_{K,j}, W_{V,j} \in\mathbb{R}^{r_Z \times r_h^{\mathrm{att}}} are learned query, key, and value projections. The head outputs are concatenated columnwise and projected as

O=Concat⁡(O1,…,OHa)WO+1qbOT∈Rq×rZ,O = \operatorname{Concat}(O_1,\ldots,O_{H_a})W_O + \mathbf{1}_q b_O^{\mathsf{T}} \in\mathbb{R}^{q \times r_Z},

where WO∈RrZ×rZW_O \in\mathbb{R}^{r_Z \times r_Z}, bO∈RrZb_O \in\mathbb{R}^{r_Z}, and 1q\mathbf{1}_q is the length-qq vector of ones. For a matrix OO, Flatten⁡(O):=(O1,:T,…,Oq,:T)T∈RqrZ\operatorname{Flatten}(O) := (O_{1,:}^{\mathsf{T}},\ldots,O_{q,:}^{\mathsf{T}})^{\mathsf{T}} \in\mathbb{R}^{qr_Z} denotes rowwise vectorization. The pooled sequence embedding is therefore

hseq=WprojFlatten⁡(O)+bproj∈RrZ,h_{\mathrm{seq}} = W_{\mathrm{proj}}\operatorname{Flatten}(O) + b_{\mathrm{proj}} \in\mathbb{R}^{r_Z},

with Wproj∈RrZ×qrZW_{\mathrm{proj}} \in\mathbb{R}^{r_Z \times qr_Z} and bproj∈RrZb_{\mathrm{proj}} \in\mathbb{R}^{r_Z}.

We augment this learned representation with two geometric features. Let u∗=arg⁡max⁡u∈{1,…,T} Zuu^* = \underset{u \in\{1,\ldots,T\}}{\arg\max}\, Z_u, and define Z~max⁡=log⁡{1+max⁡uZu}\widetilde{Z}_{\max} = \log\{1 + \max_u Z_u\} and Z~arg=(u∗−1)/T\widetilde{Z}_{\mathrm{arg}} = (u^* - 1)/T, where the subtraction by one matches the zero-based index used in the implementation. Then

hZ=[hseq,Z~max⁡,Zarg]∈RrZ+2,rc=rt+rZ+2+a.h_Z = [h_{\mathrm{seq}}, \widetilde{Z}_{\max}, Z_{\mathrm{arg}}] \in\mathbb{R}^{r_Z+2}, \qquad r_c = r_t + r_Z + 2 + a.

Given xx and ctc_t, the vector-field backbone uses Nblk=4N_{\mathrm{blk}} = 4 fixed-width residual blocks with hidden dimension rb=256r_b = 256, Adaptive Layer Normalization (AdaLN), and Swish-Gated Linear Units (SwiGLU) [36, 31]. The initial state is h(0)=Wxx+bx∈Rrbh^{(0)} = W_xx + b_x \in\mathbb{R}^{r_b}, where Wx∈Rrb×dW_x \in\mathbb{R}^{r_b \times d} and bx∈Rrbb_x \in\mathbb{R}^{r_b}. In block m∈{1,…,Nblk}m \in\{1,\ldots,N_{\mathrm{blk}}\}, the context-dependent scale and shift are

[γm(ct)βm(ct)]=Wc,mSiLU⁡(ct)+bc,m∈R2rb,\begin{bmatrix} \gamma_m(c_t) \\ \beta_m(c_t) \end{bmatrix} = W_{c,m}\operatorname{SiLU}(c_t) + b_{c,m} \in\mathbb{R}^{2r_b},

where Wc,m∈R2rb×rcW_{c,m} \in\mathbb{R}^{2r_b \times r_c} and bc,m∈R2rbb_{c,m} \in\mathbb{R}^{2r_b}. The block then computes

h~(m)=LayerNorm⁡(h(m−1))⊙{1+γm(ct)}+βm(ct),um=Wval,mh~(m)+bval,m,gm=Wgate,mh~(m)+bgate,m,h(m)=h(m−1)+Wout,mDropout⁡{um⊙SiLU⁡(gm)}+bout,m,\begin{aligned} \widetilde{h}^{(m)} &= \operatorname{LayerNorm}(h^{(m-1)}) \odot\{1+\gamma_m(c_t)\}+\beta_m(c_t), \\ u_m &= W_{\mathrm{val},m}\widetilde{h}^{(m)}+b_{\mathrm{val},m}, \qquad g_m = W_{\mathrm{gate},m}\widetilde{h}^{(m)}+b_{\mathrm{gate},m}, \\ h^{(m)} &= h^{(m-1)}+W_{\mathrm{out},m}\operatorname{Dropout}\{u_m\odot\operatorname{SiLU}(g_m)\}+b_{\mathrm{out},m}, \end{aligned}

where Wval,m,Wgate,m,Wout,m∈Rrb×rbW_{\mathrm{val},m}, W_{\mathrm{gate},m}, W_{\mathrm{out},m} \in\mathbb{R}^{r_b \times r_b}, their corresponding biases lie in Rrb\mathbb{R}^{r_b}, and ⊙\odot denotes elementwise multiplication. Here LayerNorm, Dropout, and SiLU denote the standard layer-normalization, dropout, and sigmoid-weighted linear-unit operations, respectively. Finally, LayerNorm⁡(h(Nblk))\operatorname{LayerNorm}(h^{(N_{\mathrm{blk})})} is concatenated with ctc_t and passed through dd coordinate-specific output heads g1ϕ,…,gdϕg^\phi_1,\ldots,g^\phi_d, giving

v_t^\phi(x \mid A,Z)=\left(g^\phi_1([\operatorname{LayerNorm}(h^{(N_{\mathrm{blk})}),c_t]),\ldots,g^\phi_d([\operatorname{LayerNorm}(h^{(N_{\mathrm{blk})}),c_t])\right)^\top\in\mathbb{R}^d.}}

The context projections and Wout,mW_{\mathrm{out},m} are initialized to zero, so each residual block initially acts as the identity. The output heads are separately initialized to zero, which makes the initial vector-field output vtϕv_t^\phi equal to zero and improves numerical stability early in training.

The sequence encoder is a Transformer with 4 layers, 4 attention heads, feedforward dimension 512, and dropout rate 0.1 [40]. The model is trained on the 150,000 contaminated trajectories using the conditional flow-matching objective with σmin⁡=0.1\sigma_{\min}=0.1.

Flow Matching Details

This section summarizes the general flow-matching framework used in the main text. Flow matching learns a continuous-time transport from a simple reference distribution pinitp_{\mathrm{init}} to a target distribution pdatap_{\mathrm{data}} by regressing a time-dependent vector field [24]. Let pt(x∣z)p_t(x \mid z) be a conditional probability path with endpoints p0(⋅∣z)=pinitp_0(\cdot\mid z)=p_{\mathrm{init}} and p1(⋅∣z)=N(z,σmin⁡2I)≈δzp_1(\cdot\mid z)=\mathcal{N}(z,\sigma_{\min}^2 I)\approx\delta_z. A conditional vector field vtflow(⋅∣z)v_t^{\mathrm{flow}}(\cdot\mid z) generates this path if the solution of

dXt=vtflow(Xt∣z)dt,dX_t=v_t^{\mathrm{flow}}(X_t\mid z)dt,

with X0∼p0(⋅∣z)X_0\sim p_0(\cdot\mid z) satisfies Xt∼pt(⋅∣z)X_t\sim p_t(\cdot\mid z) for each t∈[0,1]t\in[0,1].

Integrating over z∼pdataz\sim p_{\mathrm{data}} yields the marginal path

pt(x):=∫pt(x∣z)pdata(z) dz,p_t(x):=\int p_t(x\mid z)p_{\mathrm{data}}(z)\,dz,

with endpoints p0=pinitp_0=p_{\mathrm{init}} and p1=N(0,σmin⁡2I)∗pdata≈pdatap_1=\mathcal{N}(0,\sigma_{\min}^2I)*p_{\mathrm{data}}\approx p_{\mathrm{data}}. The corresponding marginal vector field is

vttarget(x)=∫vtflow(x∣z)pt(x∣z)pdata(z)pt(x) dz,v_t^{\mathrm{target}}(x)=\int v_t^{\mathrm{flow}}(x\mid z)\frac{p_t(x\mid z)p_{\mathrm{data}}(z)}{p_t(x)}\,dz,

which drives

X0∼pinit,dXt=vttarget(Xt)dt,X1∼⋅pdata.X_0\sim p_{\mathrm{init}},\qquad dX_t=v_t^{\mathrm{target}}(X_t)dt,\qquad X_1\overset{\cdot}{\sim}p_{\mathrm{data}}.

The ideal objective is

LFM(ϕ)=Et∼π, x∼pt[∥vtϕ(x)−vttarget(x)∥2],\mathcal{L}_{\mathrm{FM}}(\phi)=\mathbb{E}_{t\sim\pi,\ x\sim p_t}\left[\left\|v_t^\phi(x)-v_t^{\mathrm{target}}(x)\right\|^2\right],

but vttargetv_t^{\mathrm{target}} is often intractable. In practice, one minimizes the conditional objective

LCFM(ϕ)=Et∼π, z∼pdata, x∼pt(⋅∣z)[∥vtϕ(x)−vtflow(x∣z)∥2],\mathcal{L}_{\mathrm{CFM}}(\phi)=\mathbb{E}_{t\sim\pi,\ z\sim p_{\mathrm{data}},\ x\sim p_t(\cdot\mid z)}\left[\left\|v_t^\phi(x)-v_t^{\mathrm{flow}}(x\mid z)\right\|^2\right],

which has the same gradient target under standard regularity conditions, i.e., ∇ϕLCFM(ϕ)=\nabla_\phi\mathcal{L}_{\mathrm{CFM}}(\phi)= ∇ϕLFM(ϕ)\nabla_{\phi}\mathcal{L}_{\mathrm{FM}}(\phi) [24, 25].

After training, approximate samples are obtained by solving

X0∼pinit,dXt=vtϕ(Xt)dt,t∈[0,1],X_0 \sim p_{\mathrm{init}}, \qquad dX_t = v_t^{\phi}(X_t)dt, \qquad t \in[0,1],

and taking X1X_1.

For Gaussian probability paths,

pt(x∣z)=N(x;αtz,βt2Id),p_t(x \mid z) = \mathcal{N}(x; \alpha_t z, \beta_t^2 I_d),

with smooth schedules (αt,βt)(\alpha_t,\beta_t) satisfying α0=0\alpha_0 = 0, α1=1\alpha_1 = 1, β0=1\beta_0 = 1, and β1=σmin⁡\beta_1 = \sigma_{\min}, the conditional velocity is

vtflow(x∣z)=(α˙t−β˙tβtαt)z+β˙tβtx.v_t^{\mathrm{flow}}(x \mid z) = \left(\dot{\alpha}_t - \frac{\dot{\beta}_t}{\beta_t}\alpha_t\right)z + \frac{\dot{\beta}_t}{\beta_t}x.

In practice, training uses Monte Carlo estimates of the CFM objective. At each update, we draw a minibatch of target–condition pairs (zb,yb)(z_b,y_b) from the training data, times tb∼πt_b \sim\pi, and independent noise vectors ϵb∼N(0,Id)\epsilon_b \sim\mathcal{N}(0,I_d); we then form xtb=tbzb+{1−(1−σmin⁡)tb}ϵbx_{t_b} = t_b z_b + \{1 - (1-\sigma_{\min})t_b\}\epsilon_b and evaluate the closed-form target vtbflow(xtb∣zb)v_{t_b}^{\mathrm{flow}}(x_{t_b} \mid z_b). The network parameters ϕ\phi are updated by backpropagation to minimize the minibatch mean squared error between this target and vtbϕ(xtb∣yb)v_{t_b}^{\phi}(x_{t_b} \mid y_b), thereby learning the conditional vector field over the full probability path without solving an ODE during training.

For conditional generation, guided flows interpolate unconditional and conditional fields:

v~t(x∣y)=(1−wt)vttarget(x∣∅)+wtvttarget(x∣y),wt>1,\tilde{v}_t(x \mid y) = (1-w_t)v_t^{\mathrm{target}}(x \mid\varnothing) + w_t v_t^{\mathrm{target}}(x \mid y), \qquad w_t > 1,

which increases adherence to the condition yy at sampling time [42] (Ho and Salimans 2022).

WAIFW Matrix for the Age-Structured SEIR Model

We adopt a structured WAIFW matrix in which the entry between age classes ii and jj depends only on ∣i−j∣\lvert i-j\rvert, following Anderson and May [1] and Shea et al. [37]:

Wij={1,∣i−j∣=0,11/12,∣i−j∣=1,5/6,∣i−j∣=2,3/4,∣i−j∣=3,2/3,∣i−j∣≥4.W_{ij} = \begin{cases} 1, & \lvert i-j\rvert= 0, \\ 11/12, & \lvert i-j\rvert= 1, \\ 5/6, & \lvert i-j\rvert= 2, \\ 3/4, & \lvert i-j\rvert= 3, \\ 2/3, & \lvert i-j\rvert\ge4. \end{cases}

This off-diagonal decay encodes the empirical observation that contact rates are highest within an age class and decline gradually with age separation.

MCMC Specifications and Diagnostics

The conventional Bayesian calibration baselines reported in the main text were implemented in CmdStan (Carpenter et al. [8]) using the No-U-Turn Sampler (NUTS), an adaptive Hamiltonian Monte Carlo algorithm (Hoffman and Gelman [18]). For both the simulation study and the real-data application, we ran four independent chains with max_treedepth = 10 and adapt_delta = 0.85. The bounded parameter declarations listed below had no additional prior-density statements in the Stan programs and therefore induce uniform priors over their respective supports.

In both analyses, the time-series likelihood was a Gaussian process centered at the relevant forward-model trajectory, following the KOH discrepancy formulation (Kennedy and O’Hagan [19]). In the notation of the main text,

Z∣θ,τ,κ,ζ∼NT{Y(θ),Σ(θ,τ,κ,ζ)},(15)Z \mid\theta, \tau, \kappa, \zeta\sim\mathcal{N}_{T}\{Y(\theta), \boldsymbol{\Sigma}(\theta,\tau,\kappa,\zeta)\}, \tag*{(15)}

where

Σuv(θ,τ,κ,ζ)=σY2(θ)[ζI(u=v)+κexp⁡{−∣u−v∣τ}],u,v=1,…,T.(16)\Sigma_{uv}(\theta,\tau,\kappa,\zeta)=\sigma_Y^2(\theta)\left[\zeta I(u=v)+\kappa\exp\left\{-\frac{\lvert u-v\rvert}{\tau}\right\}\right],\qquad u,v=1,\ldots,T. \tag*{(16)}

Here τ\tau is the correlation range, κ\kappa is the partial sill, ζ\zeta is the nugget, and σY2(θ)\sigma_Y^2(\theta) is the temporal variance of the forward trajectory, as defined in Section 3.4.1 of the main text. The covariance hyperparameters had independent priors κ∼Inv-Gamma⁡(2,1)\kappa\sim\operatorname{Inv\text{-}Gamma}(2,1) and ζ∼Inv-Gamma⁡(2,1)\zeta\sim\operatorname{Inv\text{-}Gamma}(2,1), where the second argument is the scale parameter. The prior support for τ\tau differed between the two analyses as specified below.

Simulation Study

For the simulation study, the calibration parameters had independent priors

θ1,θ2,θ3∼indUniform⁡(0,1),\theta_1,\theta_2,\theta_3\overset{\mathrm{ind}}{\sim}\operatorname{Uniform}(0,1),

and the GP range parameter had prior

τ∼Uniform⁡(0.05×53,0.20×53)=Uniform⁡(2.65,10.60).\tau\sim\operatorname{Uniform}(0.05\times53,0.20\times53)=\operatorname{Uniform}(2.65,10.60).

The mean in equation (15) was the Gaussian density-shaped forward trajectory used in the main text,

Y(θ,s)=0.3+θ1+0.32π(θ3+0.1)exp⁡{−(s−θ2+0.5)2θ3+0.1},Y(\theta,s)=0.3+\frac{\theta_1+0.3}{\sqrt{2\pi(\theta_3+0.1)}}\exp\left\{-\frac{(s-\theta_2+0.5)^2}{\theta_3+0.1}\right\},

evaluated at T=53T=53 equally spaced values of ss on [−2,2][-2,2].

Each of the four chains was run for 100,000 iterations: the first 30,000 iterations were discarded as warmup and the remaining 70,000 iterations were retained for posterior inference. Thus, the four chains contributed 280,000 post-warmup draws in total. Figure S1 shows the MCMC diagnostics for a single synthetic observation data: Specifically, the chain-specific marginal posterior densities and post-warmup traces for the three forward-model parameters and the three GP covariance hyperparameters. The density curves from the four chains largely overlap, and the traces show no visually apparent chain-specific drift.

MCMC diagnostics for the simulation study

Figure S1. MCMC diagnostics for the simulation study. The left column displays chain-specific marginal posterior density estimates, and the right column displays the corresponding post-warmup traces for θ1\theta_1, θ2\theta_2, θ3\theta_3, τ\tau, κ\kappa, and ζ\zeta. Each color represents one of the four NUTS chains.

Real-Data Application

For the measles application, the six forward-model parameters had independent uniform priors induced by the following bounds:

β∼Uniform⁡(0.5×10−6,2.5×10−6),N0∼Uniform⁡(1.4×105,3.0×105),fE∼Uniform⁡(10−4,10−3),ash∼Uniform⁡(0.8,1.4),art∼Uniform⁡(0.09,0.39),t0∼Uniform⁡(0.1,10).(17)\begin{aligned} \beta&\sim\operatorname{Uniform}(0.5 \times10^{-6}, 2.5 \times10^{-6}), \qquad N_0 \sim\operatorname{Uniform}(1.4 \times10^5, 3.0 \times10^5),\\ f_E &\sim\operatorname{Uniform}(10^{-4}, 10^{-3}), \qquad a_{\mathrm{sh}} \sim\operatorname{Uniform}(0.8, 1.4),\\ a_{\mathrm{rt}} &\sim\operatorname{Uniform}(0.09, 0.39), \qquad t_0 \sim\operatorname{Uniform}(0.1, 10). \tag*{(17)} \end{aligned}

For this analysis, the GP range parameter had prior

τ∼Uniform⁡(0.1×53,0.3×53)=Uniform⁡(5.3,15.9).\tau\sim\operatorname{Uniform}(0.1 \times53, 0.3 \times53) = \operatorname{Uniform}(5.3, 15.9).

The weekly infected-population time series followed equations (S1)–(S2), with the deterministic age-structured SEIR trajectory replacing the synthetic Gaussian density-shaped mean. The age-proportion observations supplied a second likelihood component. Using the age proportions obtained from the SEIR trajectory (A(θ)A(\theta)), we defined the likelihood of the observed proportions Aobs=(0.42,0.30,0.28)TA_{\mathrm{obs}} = (0.42, 0.30, 0.28)^{\mathsf{T}} as

Aobs∣θ∼Dirichlet⁡{e2A(θ)},e2=300.(18)A_{\mathrm{obs}} \mid\theta\sim\operatorname{Dirichlet}\{e_2 A(\theta)\}, \qquad e_2 = 300. \tag*{(18)}

The time-series and age-proportion likelihood contributions were conditionally independent given the forward-model parameters and GP hyperparameters.

Each of the four chains was run for 10,000 iterations, with 3,000 warmup iterations followed by 7,000 retained draws. The combined post-warmup sample therefore contained 28,000 draws. Figure S2 reports MCMC diagnostics for the observed measles outbreak data: Specifically, the chain-specific marginal posterior densities and post-warmup traces for the six SEIR parameters and the three GP covariance hyperparameters. The four density curves show substantial overlap, and the retained traces show no obvious long-term chain-specific drift.

MCMC diagnostics for the real-data application

Figure S2. MCMC diagnostics for the real-data application. The left column displays chain-specific marginal posterior density estimates, and the right column displays the corresponding post-warmup traces for β\beta, N0N_0, fEf_E, asha_{sh}, arta_{rt}, t0t_0, τ\tau, κ\kappa, and ζ\zeta. Each color represents one of the four NUTS chains.

Ablation Study for Parameter-Set Selection

We conducted an ablation study to examine the empirical behavior of the prediction intervals when the conformal risk control step is removed. In the full procedure, the retention threshold is selected by CRC [3] to control the expected miscoverage loss. In this ablation, we instead bypass this calibration step and retain a fixed proportion of candidate draws. Specifically, we retain the top 90% of candidate draws according to two simple, non-conformal selection rules. The radius rule retains the draws with the smallest Euclidean norm of the latent sample, whereas the posterior rule retains the draws with the largest posterior density values.

After this uncalibrated filtering step, all remaining components of the pipeline are kept unchanged. The retained draws are propagated through the simulator, transformed back to the original data scale, clipped to the parameter bounds, and evaluated using the same fixed-rank empirical miscoverage loss as in the main analysis. Thus, this ablation isolates the effect of replacing the CRC-selected parameter set with a fixed 90% parameter set. The resulting miscoverage values should therefore be interpreted as empirical diagnostics, not as conformal risk-controlled guarantees.

Data settingFiltering ruleMiscoverageRMSEAvg. interval length
Simulated dataRadius0.00000.01190.0597
Simulated dataPosterior0.00000.01190.0597
Real dataRadius0.00001041.646578.77
Real dataPosterior0.00001077.626078.54

Table S1. Ablation study without conformal risk control. Instead of selecting the retention threshold by CRC, each rule retains a fixed 90% of candidate draws. RMSE is computed between the observed trajectory and the median of the model outputs, and the average interval length is averaged over time.

For the simulated data, both uncalibrated filtering rules attain zero empirical miscoverage, with nearly identical RMSE and average interval length. A similar pattern appears in the real-data analysis: both rules attain zero empirical miscoverage, with similar RMSE and average interval length.

These results suggest that, in these particular examples, retaining a fixed 90% of candidate draws produces intervals with empirical miscoverage similar across the two non-conformal filtering criteria. However, unlike the CRC-based procedure, this fixed-retention ablation does not adapt the threshold to satisfy a prescribed risk level and does not provide the finite-sample risk guarantee. The ablation therefore supports the role of CRC as the component that converts a reasonable parameter-set construction into a calibrated prediction procedure, rather than merely serving as another heuristic filtering rule.

Radius-based filtering

(a) Radius-based filtering.

Posterior-based filtering

(b) Posterior-based filtering.

Figure S3. Ablation results for the simulated-data experiment. Both panels retain 90% of the candidate draws and display the median model output, the 90% fixed-rank prediction interval, and the synthetic observation.

Radius-based filtering

(a) Radius-based filtering.

Posterior-based filtering

(b) Posterior-based filtering.

Figure S4. Ablation results for the real-data experiment. Both panels retain 90% of the candidate draws and display the median model output, the 90% fixed-rank prediction interval, and the observed data.

References

  1. [1]Anderson, R. M. and R. M. May (1985). Age-related changes in the rate of disease transmission: Implications for the design of vaccination programmes. The Journal of Hygiene 94(3), 365–436.pubmed.ncbi.nlm.nih.gov/4008922
  2. [2]Angelopoulos, A. N. and S. Bates (2023). Conformal prediction: A gentle introduction. Foundations and Trends in Machine Learning 16(4), 494–591.
  3. [3]Angelopoulos, A. N., S. Bates, A. Fisch, L. Lei, and T. Schuster (2024). Conformal risk control. In The Twelfth International Conference on Learning Representations.
  4. [4]Bayarri, M., J. Berger, J. Cafeo, G. Garcia-Donato, F. Liu, J. Palomo, R. Parthasarathy, R. Paulo, J. Sacks, and D. Walsh (2007). Computer model validation with functional output. The Annals of Statistics 35(5), 1874–1906.DOI
  5. [5]Bhatnagar, S., W. Chang, S. Kim, and J. Wang (2022). Computer model calibration with time series data using deep learning and quantile regression. SIAM/ASA Journal on Uncertainty Quantification 10(1), 1–26.arxiv.org/abs/2008.13066
  6. [6]Brent, R. P. (1973). Algorithms for Minimization without Derivatives. Prentice-Hall.DOI
  7. [7]Brynjarsdóttir, J. and A. O’Hagan (2014). Learning about physical parameters: The importance of model discrepancy. Inverse Problems 30(11), 114007.DOI
  8. [8]Carpenter, B., A. Gelman, M. D. Hoffman, D. Lee, B. Goodrich, M. Betancourt, M. Brubaker, J. Guo, P. Li, and A. Riddell (2017). Stan: A probabilistic programming language. Journal of Statistical Software 76(1), 1–32.
  9. [9]Chen, R. T. Q., Y. Rubanova, J. Bettencourt, and D. K. Duvenaud (2018). Neural ordinary differential equations. In Advances in Neural Information Processing Systems, Volume 31, pp. 6571–6583.arxiv.org/abs/1806.07366
  10. [10]Cranmer, K., J. Brehmer, and G. Louppe (2020). The frontier of simulation-based inference. Proceedings of the National Academy of Sciences 117(48), 30055–30062.DOI
  11. [11]Dax, M., J. Wildberger, S. Buchholz, S. R. Green, J. H. Macke, and B. Schölkopf (2023). Flow matching for scalable simulation-based inference. In Advances in Neural Information Processing Systems, Volume 36.arxiv.org/abs/2305.17161
  12. [12]Greenberg, D., M. Nonnenmacher, and J. Macke (2019). Automatic posterior transformation for likelihood-free inference. In Proceedings of the 36th International Conference on Machine Learning, Volume 97 of Proceedings of Machine Learning Research, pp. 2404–2414. PMLR.arxiv.org/abs/1905.07488
  13. [13]Gretton, A., K. M. Borgwardt, M. J. Rasch, B. Schölkopf, A. Smola, and K.-R. Müller (2012). A kernel two-sample test. Journal of Machine Learning Research 13, 723–773.DOI
  14. [14]Guan, L. (2023). Localized conformal prediction: A generalized inference framework for conformal prediction. Biometrika 110(1), 33–50.arxiv.org/abs/2106.08460
  15. [15]Hermans, J., A. Delaunoy, F. Rozet, A. Wehenkel, V. Begy, and G. Louppe (2022). A crisis in simulation-based inference? beware, your posterior approximations can be unfaithful. Transactions on Machine Learning Research.arxiv.org/abs/2110.06581
  16. [16]Higdon, D., J. Gattiker, B. Williams, and M. Rightley (2008). Computer model calibration using high-dimensional output. Journal of the American Statistical Association 103(482), 570–583.DOI
  17. [18]Hoffman, M. D. and A. Gelman (2014). The No-U-Turn sampler: Adaptively setting path lengths in Hamiltonian Monte Carlo. Journal of Machine Learning Research 15(47), 1593–1623.DOI
  18. [19]Kennedy, M. C. and A. O’Hagan (2001). Bayesian calibration of computer models. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 63(3), 425–464.DOI
  19. [20]Kynkäänniemi, T., M. Aittala, T. Karras, S. Laine, T. Aila, and J. Lehtinen (2024). Applying guidance in a limited interval improves sample and distribution quality in diffusion models. In Advances in Neural Information Processing Systems, Volume 37.arxiv.org/abs/2404.07724
  20. [21]Lee, J., Y. Lee, J. Kim, A. Kosiorek, S. Choi, and Y. W. Teh (2019). Set Transformer: A framework for attention-based permutation-invariant neural networks. In Proceedings of the 36th International Conference on Machine Learning, Volume 97 of Proceedings of Machine Learning Research, pp. 3744–3753. PMLR.arxiv.org/abs/1810.00825
  21. [22]Lee, J., C. Xu, and Y. Xie (2026). Flow-based conformal prediction for multi-dimensional time series. In The Fourteenth International Conference on Learning Representations.arxiv.org/abs/2502.05709
  22. [23]Lim, D., K. T. Ko, H. Hong, H. Lee, B. Choi, W. Chang, S. Choi, and J. K. Kim (2025). A history-dependent approach for accurate initial condition estimation in epidemic models. PLOS Computational Biology 21(9), e1013438.DOI
  23. [24]Lipman, Y., R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le (2023). Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations.arxiv.org/abs/2210.02747
  24. [25]Lipman, Y., M. Havasi, P. Holderrieth, N. Shaul, M. Le, B. Karrer, R. T. Q. Chen, D. Lopez-Paz, H. Ben-Hamu, and I. Gat (2024). Flow matching guide and code. arXiv preprint arXiv:2412.06264.arxiv.org/abs/2412.06264
  25. [26]Lueckmann, J.-M., J. Boelts, D. S. Greenberg, P. J. Gonçalves, and J. H. Macke (2021). Benchmarking simulation-based inference. In Proceedings of the 24th International Conference on Artificial Intelligence and Statistics, Volume 130 of Proceedings of Machine Learning Research, pp. 343–351. PMLR.arxiv.org/abs/2101.04653
  26. [27]McKay, M. D., R. J. Beckman, and W. J. Conover (1979). A comparison of three methods for selecting values of input variables in the analysis of output from a computer code. Technometrics 21(2), 239–245.
  27. [28]Minetti, A., M. Kagoli, A. Katsulukuta, H. Huerga, A. Featherstone, H. Chiotcha, D. Noel, C. Bopp, L. Sury, R. Fricke, M. Iscla, N. Hurtado, T. Ducomble, S. Nicholas, S. Kabulazi, R. F. Grais, and F. J. Luquero (2013). Lessons and challenges for measles control from unexpected large outbreak, Malawi. Emerging Infectious Diseases 19(2), 202–209.DOI
  28. [29]Papamakarios, G. and I. Murray (2016). Fast ε-free inference of simulation models with Bayesian conditional density estimation. In Advances in Neural Information Processing Systems, Volume 29, pp. 1028–1036.arxiv.org/abs/1605.06376
  29. [30]Papamakarios, G., D. C. Sterratt, and I. Murray (2019). Sequential neural likelihood: Fast likelihood-free inference with autoregressive flows. In Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, Volume 89 of Proceedings of Machine Learning Research, pp. 837–848. PMLR.arxiv.org/abs/1805.07226
  30. [31]Peebles, W. and S. Xie (2023). Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 4195–4205.arxiv.org/abs/2212.09748
  31. [32]Plumlee, M. (2017). Bayesian calibration of inexact computer models. Journal of the American Statistical Association 112(519), 1274–1285.DOI
  32. [33]Plumlee, M. (2019). Computer model calibration with confidence and consistency. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 81(3), 519–545.DOI
  33. [34]Rahimi, A. and B. Recht (2007). Random features for large-scale kernel machines. In Advances in Neural Information Processing Systems, Volume 20, pp. 1177–1184.
  34. [35]Romano, Y., E. Patterson, and E. Candes (2019). Conformalized quantile regression. In Advances in Neural Information Processing Systems, Volume 32.arxiv.org/abs/1905.03222
  35. [36]Shazeer, N. (2020). GLU variants improve transformer. arXiv preprint arXiv:2002.05202.arxiv.org/abs/2002.05202
  36. [37]Shea, K., M. J. Tildesley, M. C. Runge, C. J. Fonnesbeck, and M. J. Ferrari (2014). Adaptive management and the value of information: Learning via intervention in epidemiology. PLOS Biology 12(10), e1001970.DOI
  37. [38]Syring, N. and R. Martin (2019). Calibrating general posterior credible regions. Biometrika 106(2), 479–486.DOI
  38. [39]Tuo, R. and C. J. Wu (2015). Efficient calibration for imperfect computer models. The Annals of Statistics 43(6), 2331–2352.
  39. [40]Vaswani, A., N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin (2017). Attention is all you need. In Advances in Neural Information Processing Systems, Volume 30.arxiv.org/abs/1706.03762
  40. [41]Vovk, V., A. Gammerman, and G. Shafer (2005). Algorithmic Learning in a Random World. Springer.
  41. [42]Zheng, Q., M. Le, N. Shaul, Y. Lipman, A. Grover, and R. T. Q. Chen (2023). Guided flows for generative modeling and decision making. arXiv preprint arXiv:2311.13443.arxiv.org/abs/2311.13443

Paper details

vlnc:W_jpcpcj79
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

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