src/model.py

"""Tiny multi-output PyTorch surrogate with embedded normalization."""

from __future__ import annotations

import torch


class PumpSurrogate(torch.nn.Module):
    def __init__(
        self,
        input_mean: torch.Tensor,
        input_std: torch.Tensor,
        output_mean: torch.Tensor,
        output_std: torch.Tensor,
    ) -> None:
        super().__init__()
        self.register_buffer("input_mean", input_mean.float())
        self.register_buffer("input_std", input_std.float())
        self.register_buffer("output_mean", output_mean.float())
        self.register_buffer("output_std", output_std.float())
        self.core = torch.nn.Sequential(
            torch.nn.Linear(6, 32),
            torch.nn.ReLU(),
            torch.nn.Linear(32, 32),
            torch.nn.ReLU(),
            torch.nn.Linear(32, 6),
        )

    def forward(self, features: torch.Tensor) -> torch.Tensor:
        normalized = (features - self.input_mean) / self.input_std
        prediction = self.core(normalized)
        return prediction * self.output_std + self.output_mean