1、同步代理模型与PSO采样范围

ML/nmWTAI-ML/configs/data_gen_family_random_v2_q.yaml

@@ -43,11 +43,11 @@ params:  # 储层与井筒物理参数采样设置
   active_param_names: ["k", "skin", "wellboreC", "phi", "h"]  # 参与采样变化的物理参数列表
   log_params: ["k", "wellboreC", "h"]  # 采样时采用对数尺度的物理参数
   ranges:  # 各物理参数采样范围
-    k: [1.0e-4, 100.0]  # 渗透率采样范围
+    k: [1.0e-3, 10.0]  # 渗透率采样范围
     skin: [-10.0, 10.0]  # 表皮系数采样范围
     wellboreC: [1.0e-4, 2.0]  # 井筒储集系数采样范围
     phi: [1.0e-2, 0.50]  # 孔隙度采样范围
-    h: [2.0, 100.0]  # 储层厚度采样范围
+    h: [2.0, 50.0]  # 储层厚度采样范围
     Cf: [1.0e-6, 5.0e-3]  # 岩石压缩系数采样范围
   fixed_params:  # 固定不参与采样的物理参数
     Cf:
@@ -59,9 +59,9 @@ params:  # 储层与井筒物理参数采样设置
     strategies:  # 按参数分别定义定向采样策略
       k:
         components:  # 分段采样组件列表
-          - {prob: 0.55, range: [1.0e-4, 1.0e-1], scale: "log"}  # 定向采样组件：prob 为概率，range 为范围，scale 为采样尺度
+          - {prob: 0.45, range: [1.0e-3, 1.0e-1], scale: "log"}  # 定向采样组件：prob 为概率，range 为范围，scale 为采样尺度
-          - {prob: 0.30, range: [1.0e-1, 1.0], scale: "log"}  # 定向采样组件：prob 为概率，range 为范围，scale 为采样尺度
+          - {prob: 0.35, range: [1.0e-1, 1.0], scale: "log"}  # 定向采样组件：prob 为概率，range 为范围，scale 为采样尺度
-          - {prob: 0.15, range: [1.0, 100.0], scale: "log"}  # 定向采样组件：prob 为概率，range 为范围，scale 为采样尺度
+          - {prob: 0.20, range: [1.0, 10.0], scale: "log"}  # 定向采样组件：prob 为概率，range 为范围，scale 为采样尺度
       skin:
         components:  # 分段采样组件列表
           - {prob: 0.35, range: [-10.0, -3.0], scale: "linear"}  # 定向采样组件：prob 为概率，range 为范围，scale 为采样尺度

ML/nmWTAI-ML/scripts/train_time_conditioned.py

@@ -30,15 +30,13 @@ def main() -> None:
     parser.add_argument("--no-schedule", action="store_true")
     parser.add_argument(
         "--sample-weight-mode",
-        choices=["none", "pso_domain", "pso_domain_risk"],
+        choices=["none", "risk_region"],
         default="none",
-        help="Optional sample weighting; default keeps the original unweighted training behavior",
+        help="Optional risk-region sample weighting; default keeps the original unweighted training behavior",
     )
-    parser.add_argument("--pso-outside-weight", type=float, default=0.5)
-    parser.add_argument("--pso-inside-weight", type=float, default=1.0)
     parser.add_argument("--risk-weight", type=float, default=2.5)
     parser.add_argument("--skin-lt-minus8-weight", type=float, default=3.5)
-    parser.add_argument("--sample-weight-min", type=float, default=0.25)
+    parser.add_argument("--sample-weight-min", type=float, default=1.0)
     parser.add_argument("--sample-weight-max", type=float, default=4.0)
     args = parser.parse_args()
 
@@ -67,8 +65,6 @@ def main() -> None:
         huber_beta=float(args.huber_beta),
         use_schedule=not bool(args.no_schedule),
         sample_weight_mode=str(args.sample_weight_mode),
-        pso_outside_weight=float(args.pso_outside_weight),
-        pso_inside_weight=float(args.pso_inside_weight),
         risk_weight=float(args.risk_weight),
         skin_lt_minus8_weight=float(args.skin_lt_minus8_weight),
         sample_weight_min=float(args.sample_weight_min),

ML/nmWTAI-ML/src/training/train_time_conditioned.py

@@ -76,11 +76,9 @@ class TimeConditionedTrainConfig:
     huber_beta: float = 0.05
     use_schedule: bool = True
     sample_weight_mode: str = "none"
-    pso_outside_weight: float = 0.5
-    pso_inside_weight: float = 1.0
     risk_weight: float = 2.5
     skin_lt_minus8_weight: float = 3.5
-    sample_weight_min: float = 0.25
+    sample_weight_min: float = 1.0
     sample_weight_max: float = 4.0
     device: str = "cuda" if torch.cuda.is_available() else "cpu"
 
@@ -143,29 +141,15 @@ def _build_sample_weight(data: dict, cfg: TimeConditionedTrainConfig, split: str
     n = int(data[f"X_params_{split}"].shape[0])
     if mode in {"none", "off", "false"}:
         return np.ones((n,), dtype=np.float32)
-    if mode not in {"pso_domain", "pso_domain_risk"}:
+    if mode != "risk_region":
         raise ValueError(f"Unknown sample_weight_mode={cfg.sample_weight_mode!r}")
 
     params = _raw_params_from_processed_split(data, split)
-    pso_mask = (
+    weight = np.ones((n,), dtype=np.float32)
-        (params["k"] >= 0.001)
+    risk = (params["skin"] < -5.0) & (params["wellboreC"] > 0.1)
-        & (params["k"] <= 10.0)
+    skin_extreme = params["skin"] < -8.0
-        & (params["skin"] >= -10.0)
+    weight[risk] = np.maximum(weight[risk], float(cfg.risk_weight))
-        & (params["skin"] <= 10.0)
+    weight[skin_extreme] = np.maximum(weight[skin_extreme], float(cfg.skin_lt_minus8_weight))
-        & (params["wellboreC"] >= 1.0e-4)
-        & (params["wellboreC"] <= 2.0)
-        & (params["phi"] >= 0.01)
-        & (params["phi"] <= 0.5)
-        & (params["h"] >= 2.0)
-        & (params["h"] <= 50.0)
-    )
-
-    weight = np.where(pso_mask, float(cfg.pso_inside_weight), float(cfg.pso_outside_weight)).astype(np.float32)
-    if mode == "pso_domain_risk":
-        risk = pso_mask & (params["skin"] < -5.0) & (params["wellboreC"] > 0.1)
-        skin_extreme = pso_mask & (params["skin"] < -8.0)
-        weight[risk] = np.maximum(weight[risk], float(cfg.risk_weight))
-        weight[skin_extreme] = np.maximum(weight[skin_extreme], float(cfg.skin_lt_minus8_weight))
 
     weight = np.clip(weight, float(cfg.sample_weight_min), float(cfg.sample_weight_max))
     return weight.astype(np.float32)