The goal of this project is to classify spectral data of copper-based minerals into mineral classes, mineral subgroups and minerals. Gold-labels used for classification are based on the book Lehrbuch der Mineralogie (Rösler, 1985). The dataset consists of measurements taken using laser-induced breakdown spectroscopy (LIBS) of 100 distinct minerals. Additionally, synthetic spectral data can be generated using the code below to augment the size of the real-world data and potentially improve classification accuracy.
Access to the dataset of handheld LIBS measurements can be requested from Dr. Daniel Riebe at the Department of Physical Chemistry, university of Potsdam.
Fork and clone the repository to your drive, create a clean virtual environment (Python 3.6) with a method of your choosing, then enter the cloned directory and run
pip install -r requirements.txt
to install all packages required to run the code of this project. Then, the file config/datasets.yaml must be edited to include absolute paths to your cloned repository as well as the parent directory of the unprocessed handheld (and any possible synthetic) dataset (raw_path). The paths of the preprocessed datasets (two handheld, one synthetic) need to be added later as well.
After accessing the handheld dataset, several preparations need to be completed before it can be used with models.
First, the relevant mineral spectra must be extracted from the large collection of .csv files spread out across many nested subdirectories. Running
python src/handheld_dataset/organise_hh_dataset.py
will collate a 100-mineral dataset (HH100) and a 12-mineral subset (HH12) and organise them in separate folders each. Afterwards, they can be converted to compressed numpy (.npz) files using
python src/handheld_dataset/csvtonumpy.py
as well as split them into training and test sets while respecting measure point boundaries. The naming convention for the data files conforms to <mineral id>_<class id>_<subgroup id>_<measure point #>_<shot #>.npz
Finally, baseline correction needs to be applied to adjust for measurement noise artefacts:
python src/handheld_dataset/baselinecorrection.py
To save disk space, the unprocessed .csv data files as well as the uncorrected dataset directories can be deleted, as baseline correction creates duplicates of the datasets.
Because the handheld dataset is of limited size, synthetic data can be generated to inflate it and theoretically bolster classification performance. In our testing, 12 minerals were selected for this that were most prevalent in the handheld data and spanned 4 of the most important mineral classes. Detailed information about the gold-standard labels and atomic compositions of all minerals in both handheld and synthetic data can be found in data/mineral_info/synthetic_minerals_raw.csv.
First, the above-mentioned mineral information needs to be encoded into numeric values that serve as the labels for generated data. Running
python src/synthetic_dataset/prepare_source_list.py
will accomplish that, as well as run several checks to assert validity of the mineral data. Afterwards, the generation process can be started using
python src/synthetic_dataset/create_synthetic_np.py -p <..> -i <..>
with -p being the number of threads that can be used and -i being a manually picked worker-id prepended to finished data to guarantee unique file names. This process is an endless loop and thus needs to be stopped manually.
Place the generated synthetic data in a subdirectory named synthetic in the raw_path location defined in the config earlier. Now, run
python src/synthetic_dataset/prepare_all_synthetic.py
to split the data into train, eval and test sets (default: 70/15/15) and convert the files to .npz format.
To test whether everything is set up correctly, try running one of the models:
python src/baseline_mlp.py -e 1 -c 2 -d 1
This should start a simple model to examine a performance baseline, for 1 epoch (-e) with no repetitions on the 12-mineral subset.
For a detailed explanation of command line arguments, start any model with the -h parameter.
These models are available to test different hypotheses:
All models (except decision trees) are either using all of the baseline architecture or slightly modified parts of it. This allows them to be compared with this baseline model more easily.
python src/baseline_mlp.py -r 3 -e 10 -c 2 -d 1 -n <{0, 1, 2}>
This would run the baseline model 3 times for 10 epochs each on the HH12 dataset with minerals as the classification target using the chosen normalisation method: 0 for no normalisation, 1 for standard normal variate normalisation or 2 for min-max normalisation. Prediction accuracy results of models with a -r value > 1 will be averaged over all model repetitions. These results can then be compared with those from the other normalisation methods.
Hypothesis: LIBS device measurements are taken as 64 shots in an 8x8 grid, presumed to be aimed at the centre of minerals by the conducting researcher. Thus, classification accuracy in the centre of the grid should be higher than at the edges of the grid.
python src/64shot_mlp.py -e 10 -c 2 -d 1
This trains a baseline model for 10 epochs on the HH12 dataset with minerals as the classification target. Afterwards, the model arranges the test set's measure points into its original 8x8 grids and predicts accuracies for each shot in the grid. All intermediate heatmap results for measure points are saved to the results/ folder and the result of calculating an average across all measure point heatmaps is displayed.
There are 2 transfer learning concepts available: classic transfer learning where the final layer is replaced for fine-tuning on another dataset, and concatenated transfer learning where the output of the last dense layer of one network is concatenated to the input of another network that is fine-tuned on the secondary dataset.
For these models, no dataset can be selected with command line parameters, as they only work with a combination of the HH12 dataset and a synthetic dataset containing the same 12 minerals.
python src/concat_mlp.py -r 3 -e 10 -c 2
python src/transfer_mlp.py -r 3 -e 10 -c 2
To evaluate the influence of the generated synthetic data on the accuracy of neural network training, the following code will train a baseline model with a mixture of HH12 and the synthetic dataset: from a composition 100% handheld + 0% synthetic data to 100 % handheld + 100% synthetic data.
python src/mixture_mlp.py -r 5 -e 10 -c 2
All visualisations will be saves in data/visualisations.
Calculates and saves average synthetic and average handheld spectra for the 12 minerals:
python src/visualisation/average_spectra.py
To plot the handheld dataset:
python src/visualisation/plot_average_spectra_hh12.py
To plot the comparison of handheld and synthetic data:
python src/visualisation/plot_syn_hh_spectra.py
Visualisation of uncorrected spectrum, the baseline and the corrected spectrum
python3 src/visualisation/plot_baseline.py
python3 src/visualisation/plot_mixture_results.py
python3 src/visualisation/minmaxdataset.py
Thanks to
- Pia Brinkmann, Dr. Daniel Riebe and the Department of Physical Chemistry for providing the dataset of handheld LIBS measurements
- Federico Malerba for providing parallelisable code to generate interpolated LIBS spectra from stick plot data
- Dr. Paul Prasse for providing code to handle receiving spectral data from the NIST LIBS database