Add Neural Flow Maps paper summary

.comprehend/papers/arxiv-2312-14635/summary.json

@@ -0,0 +1,202 @@
+{
+  "title": "Fluid Simulation on Neural Flow Maps",
+  "pdf_url": "https://arxiv.org/pdf/2312.14635.pdf",
+  "tags": [
+    "flow-maps",
+    "implicit-representations",
+    "world-simulation"
+  ],
+  "slug": "arxiv-2312-14635",
+  "keywords": [
+    "Neural Flow Maps",
+    "Spatially Sparse Neural Fields",
+    "bidirectional marching",
+    "spatiotemporal velocity buffer",
+    "flow-map Jacobians",
+    "dynamic timestamp normalization",
+    "impulse advection",
+    "BFECC",
+    "sizing field",
+    "multi-resolution sparse feature grids"
+  ],
+  "problem": [
+    "Long-range flow-map advection can preserve vortical detail by avoiding repeated grid interpolation, but accurately marching a backward map requires the entire velocity history; storing that history densely makes the method's memory cost impractical.",
+    "Earlier bidirectional solvers evolve the forward map with interpolation-free Runge–Kutta marching but advect the backward map semi-Lagrangianly. This numerical asymmetry breaks the inverse-map and Jacobian consistency conditions in **4b**, producing transport error and numerical dissipation despite the conceptual symmetry shown in (5a).",
+    "Generic implicit neural representations are not a drop-in remedy: pure networks train too slowly, dense or factorized encodings do not allocate capacity to localized vortical structure, and hash grids introduce collisions where all velocity samples matter equally."
+  ],
+  "solution": [
+    "Neural Flow Maps (NFM) compresses the long-term velocity history into a continuously queryable spatiotemporal velocity buffer, then evaluates the same map and Jacobian dynamics **4a** forward and backward with a custom interleaved RK4 integrator. This mechanistic symmetry directly targets the consistency identities **4b**.",
+    "The buffer uses Spatially Sparse Neural Fields (SSNF), whose architecture is shown in (5b): coordinates query overlapping multi-resolution sparse feature grids; features are regrouped into four temporal anchors, interpolated cubically in time, and decoded by separate shallow networks into staggered MAC-grid velocity components.",
+    "SSNF explicitly activates fine cells where the velocity-gradient sizing field **4d** is large. This collision-free allocation concentrates parameters around vortices instead of relying on hash collisions, while the evolving discretization in (5c) lets capacity follow newly complex regions during simulation.",
+    "To learn an online stream of velocity frames, NFM copies the current buffer into an auxiliary teacher, trains the main buffer on the newest frame plus replayed queries from the teacher, and applies dynamic timestamp normalization so a sequence of unknown duration always occupies the interpolation interval.",
+    "At each simulation step, NFM estimates a midpoint velocity, updates the neural buffer, marches the backward map over the stored history and the forward map one step, reconstructs impulse through **4c**, applies BFECC error compensation, and projects the result to a divergence-free velocity. Periodic reinitialization balances neural fitting error against long-range marching error."
+  ],
+  "key_concepts": [
+    "Flow maps separate material transport from field storage. The forward map sends an initial material point to its current position, while the backward map retrieves its origin; their Jacobians encode local deformation. Figure (5a) anchors these four quantities geometrically.",
+    "Bidirectional marching is not merely running two unrelated advections. NFM integrates both temporal directions with the same RK4 mechanism, so round trips and composed deformation gradients remain close to identity as required by **4b**; this consistency is what permits long map horizons without the drift caused by asymmetric interpolation.",
+    "SSNF is a hybrid implicit representation rather than a monolithic MLP. Explicit multi-resolution sparse features provide local, trainable degrees of freedom, cubic temporal interpolation makes the field continuous between four anchor times, and small component-specific decoders preserve the staggered structure of a MAC grid (5b).",
+    "Domain-aware sparsity is central to the representation. The sizing field **4d** measures local velocity variation, and fine grid levels activate only where that variation exceeds their scale-dependent threshold. Consequently, storage scales with flow complexity rather than the volume of a dense space-time grid; (5c) shows this support changing as the flow evolves.",
+    "Dynamic timestamp normalization solves an online representation problem: future sequence length and frame durations are unknown. As each frame arrives, the reciprocal elapsed time rescales prior timestamps into the unit interval, allowing the same four temporal anchors to represent a growing history without assuming a fixed endpoint.",
+    "Impulse advection preserves circulation information through a covector transformation. The backward-map Jacobian transports initial impulse via **4c**, BFECC uses the forward map to estimate round-trip error, and a Poisson projection recovers incompressible velocity. The neural field supplies history to this first-principles solver; it does not emulate simulation trajectories from a pre-trained dataset."
+  ],
+  "math": [
+    {
+      "id": "4a",
+      "label": "bidirectional flow-map and Jacobian dynamics",
+      "latex": "\\frac{\\partial \\boldsymbol{\\phi}(\\mathbf{X},\\tau)}{\\partial \\tau}=\\mathbf{u}(\\boldsymbol{\\phi}(\\mathbf{X},\\tau),\\tau),\\quad \\boldsymbol{\\phi}(\\mathbf{X},0)=\\mathbf{X},\\qquad \\frac{D\\mathbf{F}}{D\\tau}=\\nabla\\mathbf{u}\\,\\mathbf{F};\\qquad \\frac{\\partial \\boldsymbol{\\psi}(\\mathbf{x},\\tau)}{\\partial \\tau}=\\mathbf{u}(\\boldsymbol{\\psi}(\\mathbf{x},\\tau),\\tau),\\quad \\boldsymbol{\\psi}(\\mathbf{x},t)=\\mathbf{x},\\qquad \\frac{D\\mathbf{T}}{D\\tau}=-\\mathbf{T}\\nabla\\mathbf{u}",
+      "variables": [
+        {
+          "symbol": "\\mathbf{X}",
+          "meaning": "material-point position at the beginning of a map interval"
+        },
+        {
+          "symbol": "\\mathbf{x}",
+          "meaning": "position of the same material point at the end of the interval"
+        },
+        {
+          "symbol": "\\boldsymbol{\\phi}",
+          "meaning": "forward flow map, integrated from the initial condition"
+        },
+        {
+          "symbol": "\\boldsymbol{\\psi}",
+          "meaning": "backward flow map, integrated backward from its terminal condition"
+        },
+        {
+          "symbol": "\\mathbf{u}",
+          "meaning": "continuous spatiotemporal velocity field queried from SSNF"
+        },
+        {
+          "symbol": "\\mathbf{F}=\\partial\\boldsymbol{\\phi}/\\partial\\mathbf{X}",
+          "meaning": "forward-map deformation gradient"
+        },
+        {
+          "symbol": "\\mathbf{T}=\\partial\\boldsymbol{\\psi}/\\partial\\mathbf{x}",
+          "meaning": "backward-map deformation gradient"
+        }
+      ]
+    },
+    {
+      "id": "4b",
+      "label": "perfect bidirectional consistency",
+      "latex": "\\boldsymbol{\\psi}\\circ\\boldsymbol{\\phi}(\\mathbf{X})=\\mathbf{X},\\qquad \\boldsymbol{\\phi}\\circ\\boldsymbol{\\psi}(\\mathbf{x})=\\mathbf{x},\\qquad \\mathbf{F}\\mathbf{T}=\\mathbf{T}\\mathbf{F}=\\mathbf{I}",
+      "variables": [
+        {
+          "symbol": "\\circ",
+          "meaning": "composition of backward and forward flow maps"
+        },
+        {
+          "symbol": "\\mathbf{I}",
+          "meaning": "identity matrix; the target composed deformation"
+        }
+      ]
+    },
+    {
+      "id": "4c",
+      "label": "impulse transport and incompressible projection",
+      "latex": "\\mathbf{m}(\\mathbf{x},t)=\\mathbf{T}^{\\mathrm{T}}\\mathbf{m}(\\boldsymbol{\\psi}(\\mathbf{x}),0),\\qquad \\nabla^2\\varphi=\\nabla\\cdot\\mathbf{m},\\qquad \\mathbf{u}=\\mathbf{m}-\\nabla\\varphi",
+      "variables": [
+        {
+          "symbol": "\\mathbf{m}",
+          "meaning": "fluid impulse transported from the initial reference frame"
+        },
+        {
+          "symbol": "\\mathbf{T}^{\\mathrm{T}}",
+          "meaning": "transpose of the backward-map Jacobian, which transforms impulse as a covector"
+        },
+        {
+          "symbol": "\\varphi",
+          "meaning": "scalar potential obtained from the Poisson equation"
+        },
+        {
+          "symbol": "\\mathbf{u}",
+          "meaning": "divergence-free velocity reconstructed from impulse"
+        }
+      ]
+    },
+    {
+      "id": "4d",
+      "label": "velocity-gradient sizing and sparse-cell activation",
+      "latex": "S(\\mathbf{x},t)=\\lVert\\nabla\\mathbf{u}(\\mathbf{x},t)\\rVert_{\\mathrm{F}}=\\left(\\sum_{i=1}^{d}\\sum_{j=1}^{d}\\left|\\frac{\\partial u_i}{\\partial x_j}\\right|^2\\right)^{1/2},\\qquad S_{\\max,l}>\\frac{\\sigma}{\\Delta x_l}\\ \\Longrightarrow\\ \\text{activate level }l",
+      "variables": [
+        {
+          "symbol": "S",
+          "meaning": "sizing field used to allocate SSNF feature-grid capacity"
+        },
+        {
+          "symbol": "\\nabla\\mathbf{u}",
+          "meaning": "velocity Jacobian, whose Frobenius norm measures local flow variation"
+        },
+        {
+          "symbol": "d",
+          "meaning": "number of spatial dimensions"
+        },
+        {
+          "symbol": "S_{\\max,l}",
+          "meaning": "maximum sizing value over the voxels controlled by a level-l cell"
+        },
+        {
+          "symbol": "\\sigma",
+          "meaning": "cell-activation threshold hyperparameter"
+        },
+        {
+          "symbol": "\\Delta x_l",
+          "meaning": "spatial grid spacing at resolution level l"
+        }
+      ]
+    }
+  ],
+  "visuals": [
+    {
+      "id": "5a",
+      "caption": "Forward and backward flow maps with their Jacobians",
+      "type": "extract",
+      "description": "Conceptual geometry of a material point traveling from its initial position to its current position and back, with the forward and backward Jacobians attached to their respective reference frames.",
+      "refs": [
+        "4a",
+        "4b"
+      ],
+      "page": 4,
+      "figure_number": 2,
+      "xref": null,
+      "clip": null,
+      "mermaid_source": null,
+      "manim_scene_path": null,
+      "manim_scene_class": null,
+      "asset_filename": "arxiv-2312-14635-5a.png"
+    },
+    {
+      "id": "5b",
+      "caption": "Spatially Sparse Neural Fields architecture",
+      "type": "extract",
+      "description": "End-to-end SSNF query path from spatial interpolation on overlapping sparse grid levels, through four-anchor temporal interpolation, to shallow component-specific decoders and a reconstructed staggered velocity field.",
+      "refs": [
+        "4d"
+      ],
+      "page": 6,
+      "figure_number": 4,
+      "xref": null,
+      "clip": null,
+      "mermaid_source": null,
+      "manim_scene_path": null,
+      "manim_scene_class": null,
+      "asset_filename": "arxiv-2312-14635-5b.png"
+    },
+    {
+      "id": "5c",
+      "caption": "Adaptive evolution of the neural buffer's sparse spatial discretization",
+      "type": "extract",
+      "description": "Sequence showing how SSNF's active multi-resolution feature-grid support grows and changes with the developing flow, illustrating domain-aware allocation during online simulation.",
+      "refs": [
+        "4d"
+      ],
+      "page": 16,
+      "figure_number": 19,
+      "xref": null,
+      "clip": null,
+      "mermaid_source": null,
+      "manim_scene_path": null,
+      "manim_scene_class": null,
+      "asset_filename": "arxiv-2312-14635-5c.png"
+    }
+  ]
+}