score-based adam dsm

.cache/known_methods_snapshot.json

@@ -176,10 +176,24 @@
         "score-based adam (gfdt, kgmm) (extended window)",
         "score-based adam (gfdt, kgmm) (parallel branches)"
       ]
+    },
+    "adam-dsm": {
+      "abbreviation": "ADAM-DSM",
+      "Method": "Score-based Adam (DSM)",
+      "parallelism": "serial",
+      "update_type": "gradient",
+      "method_goal": "optimization",
+      "emulator_use": "none",
+      "aliases": [
+        "adam-dsm",
+        "score-based adam (gfdt, dsm) (extended window)",
+        "score-based adam (gfdt, dsm) (parallel branches)"
+      ]
     }
   },
   "observed_methods": [
     "adam",
+    "adam-dsm",
     "adam-kgmm",
     "etki",
     "iekf",

CLAUDE.md

@@ -77,6 +77,7 @@ calibration_benchmark/
 │   ├── opt_kgmm_score_lm/       Score-based LM (KGMM score estimator) results (per-benchmark files)
 │   ├── opt_dsm_score_lm/        Score-based LM (DSM score estimator) results (per-benchmark files)
 │   ├── opt_kgmm_score_adam/     Score-based ADAM (KGMM score estimator) results (per-benchmark files)
+│   ├── opt_dsm_score_adam/      Score-based ADAM (DSM score estimator) results (per-benchmark files)
 │   ├── ces-eki-dmc_results/     CES-EKI-DMC ensemble-results files (UQ source)
 │   ├── gboed-uq_results/        GBOED ensemble-results files (UQ source)
 │   ├── gnki-uq_results/         IEKF (a.k.a GNKI) ensemble-results files (UQ source)
@@ -200,6 +201,7 @@ A quick taxonomy tree is rendered on the home page (`src/streamlit_app.py`).
 | `lm-kgmm` | LM-KGMM | serial | gradient | optimization | none | — |
 | `lm-dsm` | LM-DSM | serial | gradient | optimization | none | — |
 | `adam-kgmm` | ADAM-KGMM | serial | gradient | optimization | none | — |
+| `adam-dsm` | ADAM-DSM | serial | gradient | optimization | none | — |
 
 ## How to extend
 

data/opt_dsm_score_adam/leaderboard_sbadam_dsm_parallel_l63_2026-09-14.nc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c8331075ef213d07bfee8663720c7d60040fa97ca320913402dec3393e44178
+size 15752

data/opt_dsm_score_adam/leaderboard_sbadam_dsm_parallel_l96_const-force_2026-09-14.nc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46ef9d9bea3b5212e8633dbbd3841109da2de86a9015a5aebee4a32ebac787aa
+size 15752

data/opt_dsm_score_adam/leaderboard_sbadam_dsm_parallel_l96_flux-force_2026-09-14.nc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2643538d8001113ab560c50f224fb05763aafb1d78e2cd6c9b996b96da6d31d
+size 15752

data/opt_dsm_score_adam/leaderboard_sbadam_dsm_parallel_l96_vec-force_2026-09-14.nc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de41e884c7c4d1b06a93d90e5ca859fbe6f695e3a106b48f1f936de3bf6620f6
+size 15752

src/common/method_registry.py

@@ -182,6 +182,19 @@ KNOWN_METHODS = {
             "score-based adam (gfdt, kgmm) (parallel branches)",
         ],
     },
+    "adam-dsm": {
+        "abbreviation": "ADAM-DSM",
+        "Method": "Score-based Adam (DSM)",
+        "parallelism": "serial",
+        "update_type": "gradient",
+        "method_goal": "optimization",
+        "emulator_use": "none",
+        "aliases": [
+            "adam-dsm",
+            "score-based adam (gfdt, dsm) (extended window)",
+            "score-based adam (gfdt, dsm) (parallel branches)",
+        ],
+    },
 }
 
 

src/data_store.py

@@ -38,6 +38,7 @@ DATASET_FILES = {
         "opt_kgmm_score_lm/leaderboard_sblm_kgmm_parallel_l63_2026-09-02.nc",
         "opt_dsm_score_lm/leaderboard_sblm_dsm_parallel_l63_2026-08-12.nc",
         "opt_kgmm_score_adam/leaderboard_sbadam_kgmm_parallel_l63_2026-09-11.nc",
+        "opt_dsm_score_adam/leaderboard_sbadam_dsm_parallel_l63_2026-09-14.nc",
     ],
     "L96": [
         "teki_results/leaderboard_teki_l96_const-force_2026-09-10.nc",
@@ -49,6 +50,7 @@ DATASET_FILES = {
         "opt_kgmm_score_lm/leaderboard_sblm_kgmm_parallel_l96_const-force_2026-09-02.nc",
         "opt_dsm_score_lm/leaderboard_sblm_dsm_parallel_l96_const-force_2026-08-12.nc",
         "opt_kgmm_score_adam/leaderboard_sbadam_kgmm_parallel_l96_const-force_2026-09-11.nc",
+        "opt_dsm_score_adam/leaderboard_sbadam_dsm_parallel_l96_const-force_2026-09-14.nc",
     ],
     "L96_NN_FORCING": [
         "teki_results/leaderboard_teki_l96_flux-force_2026-09-10.nc",
@@ -58,6 +60,7 @@ DATASET_FILES = {
         "opt_kgmm_score_lm/leaderboard_sblm_kgmm_parallel_l96_flux-force_2026-09-02.nc",
         "opt_dsm_score_lm/leaderboard_sblm_dsm_parallel_l96_flux-force_2026-08-12.nc",
         "opt_kgmm_score_adam/leaderboard_sbadam_kgmm_parallel_l96_flux-force_2026-09-11.nc",
+        "opt_dsm_score_adam/leaderboard_sbadam_dsm_parallel_l96_flux-force_2026-09-14.nc",
     ],
     "L96_SPATIAL_FORCING": [
         "teki_results/leaderboard_teki_l96_vec-force_2026-09-10.nc",
@@ -69,6 +72,7 @@ DATASET_FILES = {
         "opt_kgmm_score_lm/leaderboard_sblm_kgmm_parallel_l96_vec-force_2026-09-02.nc",
         "opt_dsm_score_lm/leaderboard_sblm_dsm_parallel_l96_vec-force_2026-08-12.nc",
         "opt_kgmm_score_adam/leaderboard_sbadam_kgmm_parallel_l96_vec-force_2026-09-11.nc",
+        "opt_dsm_score_adam/leaderboard_sbadam_dsm_parallel_l96_vec-force_2026-09-14.nc",
     ],
 }
 

src/pages/MethodDetails.py

@@ -422,6 +422,31 @@ method_meta = {
         ],
         "notes": "**Inherits Adam's local-minimum susceptibility** (no trust region, restart, or ensemble diversity to escape a bad basin) **and LM-KGMM's cheaper-than-DSM score estimation** — the k-means/Gaussian-mixture score fit is far less costly than training a neural score network, so ADAM-KGMM stayed tractable on problems where a DSM-based Adam variant would not.",
     },
+    "ADAM-DSM": {
+        "citation": "Giorgini, Bischoff & Souza, arXiv:2509.19660, 2025",
+        "url": "https://arxiv.org/abs/2509.19660",
+        "source": "`opt_experiments/score_based_adam` — Adam with a GFDT/DSM gradient surrogate.",
+        "math": "Identical Adam + GFDT scheme as ADAM-KGMM (see that entry's update equation), but the score `s(x)=∇log p(x)` is estimated by Denoising Score Matching (DSM) instead of KGMM.",
+        "guardrails": (
+            "Same DSM training scheme as LM-DSM: a neural network is trained to denoise samples perturbed "
+            "by Gaussian noise at a single fixed scale `σ`, whose optimal denoiser equals the noised "
+            "density's score (Vincent, 2011); `σ` is fixed per benchmark rather than annealed, chosen "
+            "empirically as the smallest noise level at which the network still receives a usable learning "
+            "signal. Same Stein recalibration and collapsed-window `ForwardDiff.jl` fallback, and the same "
+            "Tukey-tapered response integral and evaluation-budget cap as LM-DSM/ADAM-KGMM. Same per-case "
+            "step size `α` (shared `β₁`/`β₂`/`ε`) as ADAM-KGMM — sized to ~0.25× the case's prior standard "
+            "deviation, since a single global value tuned for one benchmark's residual scaling left Adam "
+            "unable to traverse even the L63/L96-const prior within its iteration budget."
+        ),
+        "params": [
+            {"Parameter": "α (step size)", "Value": "Per-case, ~0.25× prior std (0.025–0.75 across benchmarks)"},
+            {"Parameter": "β₁, β₂", "Value": "0.9, 0.999"},
+            {"Parameter": "σ (DSM noise scale)", "Value": "0.2 (L63) / 0.025 (L96 variants)"},
+            {"Parameter": "tukey_alpha (response taper)", "Value": "0.75"},
+            {"Parameter": "budget_total (outer_iter × N_ens)", "Value": "500"},
+        ],
+        "notes": "**Inherits both LM-DSM's slow DSM-training cost and Adam's local-minimum susceptibility** — training the DSM score network is the binding wall-clock cost (as with LM-DSM), and Adam's lack of a trust region or ensemble diversity leaves it as prone to getting stuck in a bad basin as ADAM-KGMM.",
+    },
 }
 
 # KNOWN_METHODS is the source of truth for the abbreviation/full-name list,