Hi Guna
As per my understanding, the performance of your model degrades significantly when applied to the real NMR data. This issue is quite common in machine learning as the models trained on simulated data often do not generalize well to real-world data due to differences in noise, baseline, resolution, and unexpected artifacts.
Kindly refer to the following workarounds ,which could help in resolving the issue -
1. To make simulated data more like real NMR spectra, you can add noise and baseline distortions
noisy_data = simulated_data + 0.01 * randn(size(simulated_data));
x = linspace(0, 1, size(simulated_data, 2));
baseline = 0.01 * (x - 0.5).^2;
baseline_data = noisy_data + baseline;
2. Sometimes raw peak areas are too sensitive. PCA can reduce dimensionality and noise
[coeff, score, ~] = pca(training_data);
X_reduced = score(:, 1:10); % use first 10 principal components
3. Use 10 - 20% of real-world data to train or fine-tune the model
% Assume net is your pretrained network
net = train(net, real_inputs_subset, real_targets_subset);
I hope this helps.
