| { |
| "paper_id": "universal-neural-operators", |
| "paper_title": "Towards Universal Neural Operators through Multiphysics Pretraining", |
| "D1": [ |
| { |
| "id": "universal-neural-operators-D1-001", |
| "claim": "[model_parameter_counts] Approximate parameter counts for each model architecture: MambaFNO ≈10^7, PerceiverIO ≈10^8, FNO ≈10^6, Swin-v2 ≈10^9, CoDA-NO ≈10^8 (from Table 1, Param. column)", |
| "source": "Section 4, Table 1 (Param. column)" |
| }, |
| { |
| "id": "universal-neural-operators-D1-002", |
| "claim": "[experiment_scenarios] The experimental evaluation covers three distinct transfer learning scenarios: (1) out-of-sample parameter values — same PDEs with different coefficient ranges; (2) input function set extension — base equation pretraining with added physics terms during fine-tuning; (3) multi-physics cross-domain transfer — joint pretraining on diverse PDEs followed by fine-tuning on a different PDE class", |
| "source": "Section 4, paragraph 1" |
| }, |
| { |
| "id": "universal-neural-operators-D1-003", |
| "claim": "[model_architectures_tested] Model architectures evaluated: post-lifting (PL) MambaFNO, post-lifting (PL) LocalAttnFNO, Perceiver IO-based neural operator, Swin-v2 transformer, CoDA-NO (codomain attention neural operator); baseline: default FNO (Fourier Neural Operator)", |
| "source": "Section 4, paragraph 2" |
| }, |
| { |
| "id": "universal-neural-operators-D1-004", |
| "claim": "[datasets_used] PDE datasets used across experiments: Burgers' equation (nonlinear wave dynamics), Gray-Scott model (reaction-diffusion pattern formation), Navier-Stokes equations (incompressible flow), advection equation (PDEBench), heat equation (with convection extension), reaction-diffusion equation (PDEBench)", |
| "source": "Section 4, paragraphs 1, 3-4" |
| }, |
| { |
| "id": "universal-neural-operators-D1-005", |
| "claim": "[evaluation_metrics] Primary evaluation metric: range-normalized mean absolute error (NMAE, Eq 3); secondary metrics: mean squared error (MSE), average epoch time (seconds), and model parameter count", |
| "source": "Section 4, paragraph 2" |
| }, |
| { |
| "id": "universal-neural-operators-D1-006", |
| "claim": "[adapter_transfer_approach] Pretraining-fine-tuning via adapter-based architecture: lift and projection blocks serve as problem-specific adapters; during fine-tuning, the shared operator core (theta_F) is frozen and only new adapter parameters (theta_P_ft, theta_L_ft) are trained, reducing computational cost", |
| "source": "Section 3, Pre-training and fine-tuning paragraphs" |
| } |
| ], |
| "D2": [ |
| { |
| "id": "universal-neural-operators-D2-001", |
| "claim": "PDE Problem Formulation: L_p u(t, x) = f(t, x) on bounded domain D", |
| "source": "Section 3, paragraph 1" |
| }, |
| { |
| "id": "universal-neural-operators-D2-002", |
| "claim": "Neural Operator Integral Kernel Layer (Eq 1): F_t(x) = sigma( A_t v_t(x) + integral_D kappa_t(x, y) v_t(y) dy + b_t(x) )", |
| "source": "Section 3, paragraph 2, Equation (1)" |
| }, |
| { |
| "id": "universal-neural-operators-D2-003", |
| "claim": "Lifting Layer Mapping: L(a) = sigma( A_L * a + b_L )", |
| "source": "Section 3, paragraph 3" |
| }, |
| { |
| "id": "universal-neural-operators-D2-004", |
| "claim": "Projection Layer Mapping: tilde{u}(x) = P( v_{n_layers}(x) ) = sigma( A_P * v_{n_layers}(x) + b_P )", |
| "source": "Section 3, paragraph 3" |
| }, |
| { |
| "id": "universal-neural-operators-D2-005", |
| "claim": "Neural Operator Full Forward Composition: G_theta(a) = P circ F_{n_layers} circ ... circ F_1 circ L(a)", |
| "source": "Section 3, paragraph 3" |
| }, |
| { |
| "id": "universal-neural-operators-D2-006", |
| "claim": "Operator Core Parameters: theta_F = { A_t, b_t, theta_{k,t} : t = 1, ..., n_layers }", |
| "source": "Section 3, paragraph 2" |
| }, |
| { |
| "id": "universal-neural-operators-D2-007", |
| "claim": "Mamba SSM Latent Preconditioner (Eq 2): tilde{v}_0(x, t) = sum_{tau <= t} K_tau * v_0(x, t - tau)", |
| "source": "Section 3, paragraph 4, Equation (2)" |
| }, |
| { |
| "id": "universal-neural-operators-D2-008", |
| "claim": "Perceiver IO Cross-Attention Input Encoding: K_1 = FNO_{K_1}(X), V_1 = FNO_{V_1}(X), Q_1 = L", |
| "source": "Section 3, paragraph 6" |
| }, |
| { |
| "id": "universal-neural-operators-D2-009", |
| "claim": "Perceiver IO Self-Attention on Latent Representations: L' = SelfAttention(Q=L, K=L, V=L)", |
| "source": "Section 3, paragraph 7" |
| }, |
| { |
| "id": "universal-neural-operators-D2-010", |
| "claim": "Perceiver IO Output Cross-Attention Decoding: Output = CrossAttention(Q = FNO_Q(X'), K = FNO_K(L'), V = FNO_V(L'))", |
| "source": "Section 3, paragraph 7" |
| }, |
| { |
| "id": "universal-neural-operators-D2-011", |
| "claim": "Codomain Attention Mechanism: Output_m = sum_j softmax_j( sim(q_m, k_j) ) * v_j, sim(q_m, k_j) = dot_product between features, not samples", |
| "source": "Section 3, paragraph 8" |
| }, |
| { |
| "id": "universal-neural-operators-D2-012", |
| "claim": "NMAE Evaluation Metric (Eq 3): NMAE(theta) = (1 / |D_test|) * sum_{(a,u) in D_test} ||G_theta(a) - u||_{1,G} / (max_G u - min_G u + epsilon)", |
| "source": "Section 4, paragraph 2, Equation (3)" |
| }, |
| { |
| "id": "universal-neural-operators-D2-013", |
| "claim": "Pretraining: Joint Optimization of All Parameters: theta* = argmin_theta sum_{i=1..N} Loss_i( G_{theta_{P_i}, theta_F, theta_{L_i}}(a_i), u_i ), theta = {theta_{P_1}, ..., theta_{P_N}, theta_F, theta_{L_1}, ..., theta_{L_N}}", |
| "source": "Section 3, paragraph 10" |
| }, |
| { |
| "id": "universal-neural-operators-D2-014", |
| "claim": "Fine-Tuning: Adapter-Only Training with Frozen Core: theta_{ft}* = argmin_{theta_{P_ft}, theta_{L_ft}} Loss_ft( G_{theta_{P_ft}, theta_F^{frozen}, theta_{L_ft}}(a_ft), u_ft )", |
| "source": "Section 3, paragraph 11" |
| }, |
| { |
| "id": "universal-neural-operators-D2-015", |
| "claim": "Adapter Parameter Design Principle: N_adapter << N_core, adapter_variance << total_model_variance", |
| "source": "Section 3, paragraph 9" |
| }, |
| { |
| "id": "universal-neural-operators-D2-016", |
| "claim": "MambaFNO Architecture Variant: MambaFNO(a) = P circ F_{n_layers} circ ... circ F_1 circ M_phi circ L(a)", |
| "source": "Section 3, paragraph 4" |
| }, |
| { |
| "id": "universal-neural-operators-D2-017", |
| "claim": "PerceiverFNO Architecture Variant: PerceiverFNO(a) = P circ OutputCrossAttn circ SelfAttn circ InputCrossAttn circ L(a)", |
| "source": "Section 3, paragraphs 5-7" |
| } |
| ], |
| "D3": [ |
| { |
| "id": "universal-neural-operators-D3-001", |
| "claim": "Evaluate the transfer learning capability of neural operators when pretraining and fine-tuning share the same PDE equations but differ in coefficient parameter values, comparing pretrained models against training from scratch on Burgers' equation, Gray-Scott reaction-diffusion, and Navier-Stokes incompressible flow. Model baselines: MambaFNO, PerceiverIO, and CoDA-NO each evaluated in both pretrained (fine-tuned with frozen core) and scratch (trained from scratch on fine-tuning data only) modes; default FNO (scratch) as the kernel-integral NO baseline; Swin-v2 (pretrained+scratch, p.+s.) as the transformer baseline. Evaluation metrics: range-normalized mean absolute error (NMAE, %), mean squared error (MSE), and average epoch time (seconds). Results reported in Table 1.", |
| "source": "Section 4, Out-of-sample parameter values scenario, Table 1" |
| }, |
| { |
| "id": "universal-neural-operators-D3-002", |
| "claim": "Evaluate the adapter-based transfer approach when the fine-tuning PDE introduces additional input functions not present during pretraining — specifically adding convection to the heat equation and advection to reaction-diffusion equations — testing whether new adapters can absorb the extra input channels while reusing the frozen pretrained operator core. Model baselines: MambaFNO, PerceiverIO, and CoDA-NO each evaluated in pretrained (frozen core, new adapters only) vs. scratch (full training from scratch on the extended equation) modes; default FNO (scratch) as baseline; Swin-v2 (p.+s.) as transformer baseline. Evaluation metrics: NMAE (%), MSE, and average epoch time (seconds). Results reported in Table 2 jointly with the multi-physics scenario.", |
| "source": "Section 4, Input function set extension scenario, Table 2" |
| }, |
| { |
| "id": "universal-neural-operators-D3-003", |
| "claim": "Evaluate whether a neural operator jointly pretrained on multiple distinct physics systems (advection equation and Burgers' equation from PDEBench) can transfer learned representations to an entirely different PDE class (reaction-diffusion from PDEBench) through adapter-based fine-tuning with frozen core parameters. Model baselines: MambaFNO, PerceiverIO, and CoDA-NO each evaluated in pretrained (jointly pretrained on advection + Burgers, fine-tuned on reaction-diffusion with frozen core) vs. scratch (trained from scratch on reaction-diffusion data only) modes; default FNO (scratch) as baseline; Swin-v2 (p.+s.) as transformer baseline. Evaluation metrics: NMAE (%), MSE, and average epoch time (seconds). Results reported in Table 2 jointly with the input extension scenario.", |
| "source": "Section 4, General multi-physics learning, Table 2" |
| }, |
| { |
| "id": "universal-neural-operators-D3-004", |
| "claim": "Evaluate the large-transformer baseline comparison protocol: compare Swin-v2 (pretrained+scratch, p.+s., ~10^9 parameters, hierarchical vision transformer with shifted windows) against adapter-based neural operators (MambaFNO ~10^7, PerceiverIO ~10^8, CoDA-NO ~10^8) in transfer learning settings to determine whether scaling model size provides competitive or superior transfer capability compared to adapter-based architecture design with frozen-core fine-tuning. Dataset: all three transfer scenarios — out-of-sample (Burgers, Gray-Scott, Navier-Stokes), input extension (heat+convection, reaction-diffusion+advection), multi-physics (advection+Burgers pretraining to reaction-diffusion fine-tuning). Baselines: FNO (scratch, ~10^6) as the kernel-integral NO reference. Metrics: NMAE (%), MSE, parameter count. Results reported in Tables 1-2.", |
| "source": "Section 4, Tables 1-2" |
| }, |
| { |
| "id": "universal-neural-operators-D3-005", |
| "claim": "Evaluate computational efficiency of adapter-based transfer learning: compare average epoch time (seconds) across all three transfer scenarios for MambaFNO, PerceiverIO, and CoDA-NO in pretrained (fine-tuned, adapter-only parameter training) vs. scratch (full parameter training from scratch) modes. Measure speedup as the ratio of scratch epoch time to pretrained epoch time per model per scenario to quantify the training cost reduction from frozen-core adapter fine-tuning. Dataset: all three transfer scenarios (out-of-sample, input extension, multi-physics). Baselines: FNO (scratch) epoch time for absolute reference. Metrics: average epoch time (s), speedup ratio (scratch/pretrained). Results reported in Tables 1-2, Avg. epoch (s) column.", |
| "source": "Section 4, Tables 1-2" |
| }, |
| { |
| "id": "universal-neural-operators-D3-006", |
| "claim": "Evaluate whether the relative benefit of pretraining is architecture-dependent: compare the NMAE reduction ratio (scratch NMAE divided by pretrained NMAE) across MambaFNO (SSM-based), PerceiverIO (attention-based), and CoDA-NO (codomain-attention-based) within each of the three transfer scenarios. Determine whether certain architecture families benefit disproportionately from adapter-based pretraining and frozen-core fine-tuning, and whether architecture choice interacts with transfer scenario type. Dataset: all three transfer scenarios. Baselines: each architecture's own scratch training mode as self-baseline; FNO (scratch) as absolute reference. Metrics: NMAE reduction ratio, MSE reduction ratio. Results computed from Tables 1-2.", |
| "source": "Section 4, Tables 1-2" |
| } |
| ], |
| "D4": [ |
| { |
| "id": "universal-neural-operators-D4-001", |
| "claim": "Experiment phases: 1. Generate data for each equation with pretraining coefficient ranges (output: pretraining dataset, feeds step 2) -> 2. Pretrain: jointly optimize all adapters + shared core on all pretraining data (output: pretrained model, feeds step 4) -> 3. Generate data for each equation with fine-tuning (different) coefficient ranges (output: fine-tuning dataset, feeds steps 4-6) -> 4. Fine-tune: load pretrained core, freeze it, create new adapters, train adapters only -> 5. Scratch baseline: train same architecture from scratch on fine-tuning data -> 6. Evaluate all models on test data using NMAE, MSE, and record epoch times -> 7. Compare pretrained vs scratch for each model; compare against FNO and Swin-v2", |
| "source": "Section 4, Out-of-sample parameter values scenario, Table 1" |
| }, |
| { |
| "id": "universal-neural-operators-D4-002", |
| "claim": "Experiment phases: 1. Generate data for base equation with standard input function set (output: base equation dataset, feeds step 2) -> 2. Pretrain on base equation data with standard adapters (output: pretrained model, feeds step 5) -> 3. Generate data for extended equation with additional input functions (output: extended equation dataset, feeds steps 5-6) -> 4. Create new adapter pair: lifting with more input channels, projection unchanged -> 5. Fine-tune: load pretrained core frozen, train new adapters on extended equation data -> 6. Scratch baselines: train full architectures from scratch on extended equation -> 7. Evaluate and compare", |
| "source": "Section 4, Input function set extension scenario, Table 2" |
| }, |
| { |
| "id": "universal-neural-operators-D4-003", |
| "claim": "Experiment phases: 1. Generate/load advection equation and Burgers equation data (output: pretraining dataset, feeds step 2) -> 2. Pretrain: jointly optimize (L_adv, P_adv), (L_burg, P_burg), and theta_F on both equations (output: jointly pretrained model, feeds step 5) -> 3. Load PDEBench reaction-diffusion data (output: fine-tuning dataset, feeds steps 5-6) -> 4. Create new adapter pair for reaction-diffusion -> 5. Fine-tune: freeze theta_F, train only (L_rd, P_rd) on reaction-diffusion data -> 6. Scratch baselines: train from scratch on reaction-diffusion data -> 7. Evaluate and compare all models", |
| "source": "Section 4, General multi-physics learning, Table 2" |
| }, |
| { |
| "id": "universal-neural-operators-D4-004", |
| "claim": "Experiment phases: 1. For each experiment scenario (out-of-sample, extension, multi-physics): -> 2. Run pretraining for MambaFNO, PerceiverIO, CoDA-NO on pretraining data -> 3. Run fine-tuning for all three pretrained models on fine-tuning data (frozen core only) -> 4. Run scratch training for MambaFNO, PerceiverIO, FNO, CoDA-NO on fine-tuning data -> 5. Run Swin-v2 in its (p.+s.) mode on fine-tuning data -> 6. Evaluate all models on test data, recording NMAE, MSE, and epoch times -> 7. Report results per Table 1 (out-of-sample) and Table 2 (extension + multi-physics)", |
| "source": "Section 4, Overall experimental protocol, Tables 1-2" |
| }, |
| { |
| "id": "universal-neural-operators-D4-005", |
| "claim": "Experiment phases: Phase 1 (Pretraining): Jointly train all N adapters + shared core on N physics problems (output: pretrained model with shared core, feeds Phase 2) -> Phase 2 (Fine-tuning): Freeze core, train only new adapters on target problem (input: Phase 1 frozen core + new adapters, output: fine-tuned model) -> The Universal Neural Operators paper does not specify whether additional fine-tuning can be stacked (e.g., multi-stage adaptation across further PDEs after initial transfer)", |
| "source": "Section 3, Pre-training and fine-tuning paragraphs" |
| } |
| ] |
| } |
