Dataset and algorithm to perform Spatial implicit neural representation for structured electron-beam patterns

A README file is provided with the complete instructions.

Brief description:

This code implements a coordinate-based neural network that predicts image intensity from two inputs:

1. the spatial coordinate `(x, y)` of each pixel, and2. a vector of experimental control parameters (for example electrode biases).

The intended application is the forward modelling of structured electron-beam illumination patterns, where one wants a fast surrogate model of the mapping

`control parameters -> illumination pattern`.

A second workflow is also included: once the network has been trained, the input parameter vector can be optimized to reproduce a desired target pattern. This corresponds to an inverse-design step in control-parameter space.

## What the code does

The module is organized around five tasks.

### 1. Data loading and normalization

`load_pattern_dataset()` reads a stack of `.npy` images from disk together with the corresponding control-parameter array. Images are rescaled to the interval `[0, 1]`. Labels are standardized by subtracting the mean and dividing by the standard deviation.

### 2. Conversion to coordinate-based training data

`flatten_dataset()` converts an image stack of shape `(n_images, nx, ny)` into a flattened regression dataset where each training sample is a single pixel. For each pixel, the model receives:

- its spatial coordinate `(x, y)`, and- the control-parameter vector associated with the parent image.

The regression target is the pixel intensity.

This is the standard way to train an implicit neural representation (INR).

### 3. Model construction

`build_model()` creates a branched neural network with:

- one branch for spatial coordinates,- one branch for experimental parameters,- additive fusion of the two branches,- additional hidden layers after fusion,- a scalar output corresponding to image intensity.

Two variants are supported:

- a standard dense network with `tanh` activations,- a SIREN-based model using `SinusodialRepresentationDense` from `tf_siren`.

Please reference the provided README file for the rest of the details

完整操作说明已附于README文件中。

### 简要说明
本代码实现了一款基于坐标的神经网络，可通过两类输入预测图像像素强度：
1. 每个像素的空间坐标`(x, y)`；
2. 实验控制参数向量（例如电极偏置电压）。

本项目的目标应用为结构化电子束照明图案的前向建模，即构建从「控制参数」到「照明图案」的快速替代模型。
此外本项目还包含另一工作流程：在神经网络完成训练后，可通过优化输入参数向量来复现指定的目标图案，这对应于控制参数空间中的逆向设计步骤。

## 代码功能说明
本代码模块围绕五大任务构建。

### 1. 数据加载与标准化
`load_pattern_dataset()` 函数可从磁盘读取`.npy`格式的图像栈及其对应的控制参数数组。图像将被归一化至`[0, 1]`区间，标签则通过减去均值并除以标准差完成标准化。

### 2. 转换为基于坐标的训练数据
`flatten_dataset()` 函数可将形状为`(n_images, nx, ny)`的图像栈转换为扁平化的回归数据集，其中每个训练样本对应单个像素。对于每个像素，模型的输入包括：
- 其空间坐标`(x, y)`；
- 所属图像对应的控制参数向量。
回归任务的目标为该像素的强度值。这是训练隐式神经表示（Implicit Neural Representation, INR）的标准流程。

### 3. 模型构建
`build_model()` 函数可构建分支式神经网络，结构如下：
- 用于处理空间坐标的分支；
- 用于处理实验参数的分支；
- 两支特征的加性融合；
- 融合后增设的隐藏层；
- 对应图像强度的标量输出。

本实现支持两种网络变体：
- 采用`tanh`激活函数的标准全连接网络；
- 基于SIREN的模型，使用`tf_siren`库中的`SinusodialRepresentationDense`层。

其余细节请参阅附带的README文件。