# Cross-domain amortized inverse operator

- ID: 325
- Canonical URL: https://synthcore.org/idea/325/cross-domain-amortized-inverse-operator
- API JSON: https://synthcore.org/api/idea/325.json
- API Markdown: https://synthcore.org/api/idea/325.md
- Verification status: failed_benchmark
- Source: [arXiv:2607.01128](https://arxiv.org/abs/2607.01128)
- Category: architecture
- Solves: accuracy, sample-efficiency, scalability
- ML areas: attention, graph-nn, embedding, inference-speedup
- Math tags: geometry, functional-analysis, pde, linear-algebra
- Ratings: usefulness 7/10; difficulty 5/10; novelty 6/10

## Idea description

Use separate measurement-domain and target-domain token sets so a network can infer a field on one spatial domain from sparse observations on another in one forward pass. The same decoder can answer arbitrary target query points, avoiding an optimization loop for each inverse instance.

## Mathematical statement

The paper distinguishes the forward operator $\mathcal{S}:\mathcal{A}\to\mathcal{U}$, which maps $a=(c,q,u_b,\chi_D)$ to the full solution $u$, from the inverse setting, where the full field is unavailable and only observables are given through a measurement operator $\mathcal{M}:\mathcal{U}\to\mathcal{Y}$. Here $\mathcal{U}$ is the space of fields on the target domain and $\mathcal{Y}$ is the measurement space. The transferable construction is a learned approximate inverse $\widehat{\mathcal{I}}:\mathcal{Y}\times\mathcal{G}\to\mathcal{U}$, where $\mathcal{G}$ describes source and target geometry. Given measurement tokens $h_j=E_y([y_j,m_j,\chi_{D_s}(y_j)])$, target query tokens $t_i=E_x([x_i,\chi_{D_t}(x_i),\mathrm{geometry}(x_i)])$, and geometry tokens $g_k$, decode with cross-attention from target queries to measurements and geometry: $r_i=\sum_j\beta_{ij}V_yh_j+\sum_k\gamma_{ik}V_gg_k$, with $\beta_{ij}$ and $\gamma_{ik}$ separately normalized softmax weights. Train against the target field using a masked loss over valid target points, so the architecture naturally supports different source and target domains and sparse measurements.

## Key formulas

- $$\mathcal{M}:\mathcal{U}\to\mathcal{Y},\qquad y=\mathcal{M}(u),$$
- $$\widehat{\mathcal{I}}:\mathcal{Y}\times\mathcal{G}\to\mathcal{U},\qquad \widehat u=\widehat{\mathcal{I}}(y,\mathcal{G}),$$
- $$\beta_{ij}=\frac{\exp(q_i^{\mathsf T}k_j/\sqrt r)}{\sum_{\ell=1}^{N_s}\exp(q_i^{\mathsf T}k_\ell/\sqrt r)},\qquad \gamma_{ik}=\frac{\exp(q_i^{\mathsf T}\bar k_k/\sqrt r)}{\sum_{\ell=1}^{K}\exp(q_i^{\mathsf T}\bar k_\ell/\sqrt r)},$$
- $$\widehat u(x_i)=D\!\left(t_i,\sum_{j=1}^{N_s}\beta_{ij}V_yh_j+\sum_{k=1}^{K}\gamma_{ik}V_gg_k\right),\qquad \mathcal{L}=\frac{\sum_i\chi_{D_t}(x_i)\lVert \widehat u(x_i)-u(x_i)\rVert_2^2}{\sum_i\chi_{D_t}(x_i)}.$$

## Implementation notes

(1) Integration point: modify an encoder-decoder used for inverse problems, sparse sensing, neural fields, or sensor-to-field reconstruction. The encoder consumes observations $m_j$ at source locations $y_j$; the decoder accepts arbitrary target coordinates $x_i$, including a target domain with a different shape or resolution. Include a binary valid-domain mask and boundary descriptors in both token types. (2) Pseudocode: construct $h_j=E_y([m_j,y_j,\mathrm{source\_geometry}_j])$; construct geometry tokens $g_k$ from boundary samples, normals, signed distance, and global shape pooling; for each target query form $q_i=Q(E_x([x_i,\chi_{D_t}(x_i),\mathrm{target\_geometry}_i]))$; calculate measurement keys and values $(k_j,V_yh_j)$ and geometry keys and values $(\bar k_k,V_gg_k)$; calculate $\beta_{ij}$ and $\gamma_{ik}$ with the two softmax equations; decode $\widehat u(x_i)=D(t_i,\sum_j\beta_{ij}V_yh_j+\sum_k\gamma_{ik}V_gg_k)$. Train with the masked normalized MSE and optionally add a PDE residual only during training. (3) Estimate from data: $\mathcal M$ is represented by the actual sensor sampling pattern, while all attention weights are computed directly; geometry features can be calculated from meshes or point clouds. (4) First cheap experiment: use a 2D Poisson or Darcy inverse task where 16--64 boundary or interior sensors must reconstruct a dense field on unseen shapes. Compare against a neural field conditioned on pooled measurements, a standard cross-attention model without geometry tokens, and per-instance gradient-based inversion. Measure relative $L^2$ reconstruction error, calibration under sensor dropout, inference latency, and the number of optimization steps avoided. Success is lower error on unseen geometries and constant one-pass latency as target resolution increases, with graceful degradation under missing measurements.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: harms
- Verdict: Built a cross-domain amortized inverse operator with separate measurement tokens, arbitrary target queries, source/target geometry conditioning, and cross-attention, plus a pooled-measurement baseline. The mathematical mechanism checks passed: softmax normalization stayed below 3.4e-16 across source-count sweeps, uniform duplicated-token invariance stayed below 1.2e-16, and geometry normalization stayed below 3.4e-16. In the toy reconstruction, the idea achieved lower relative L2 error than the baseline (0.288 vs 0.424), degraded less under 50% sensor dropout (0.667 vs 0.724), and remained stable when target resolution increased from 64 to 256 points (0.286 vs 0.278).

### Mechanism check

- Verdict: Built a cross-domain amortized inverse operator with separate measurement tokens, arbitrary target queries, source/target geometry conditioning, and cross-attention, plus a pooled-measurement baseline. The mathematical mechanism checks passed: softmax normalization stayed below 3.4e-16 across source-count sweeps, uniform duplicated-token invariance stayed below 1.2e-16, and geometry normalization stayed below 3.4e-16. In the toy reconstruction, the idea achieved lower relative L2 error than the baseline (0.288 vs 0.424), degraded less under 50% sensor dropout (0.667 vs 0.724), and remained stable when target resolution increased from 64 to 256 points (0.286 vs 0.278).
- Confidence: 8/10
- Limitations: This is a synthetic smooth-field toy problem rather than a Poisson/Darcy PDE or unseen-mesh benchmark. Geometry uses compact analytic descriptors instead of boundary point clouds, the geometry branch has only one token in the implemented toy, and no per-instance gradient-based inversion or calibrated uncertainty was tested.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 0/8
- Benchmark verdict: idea worse (significant)

## Artifacts

- [bench_experiment.py](https://synthcore.org/code/112/bench_experiment.py)
- [bench_report.json](https://synthcore.org/code/112/bench_report.json)
- [experiment.py](https://synthcore.org/code/112/experiment.py)
- [report.md](https://synthcore.org/code/112/report.md)
- [report_bench_2026-08-30T145631.md](https://synthcore.org/code/112/report_bench_2026-08-30T145631.md)
- [results.json](https://synthcore.org/code/112/results.json)
- [signature_observed.json](https://synthcore.org/code/112/signature_observed.json)
- [Download all files as ZIP](https://synthcore.org/download/112)

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.