Added link to SU2 DataMiner repo f.

Author: EvertBunschoten

_tutorials/incompressible_flow/Inc_Combustion/Inc_Combustion.md

@@ -253,7 +253,7 @@ The table format supported under the option `LUT` allows for the interpolation o
 Examples of such 2D and 3D tables can be found in the `SU2` [unit tests](https://github.com/su2code/SU2/tree/master/UnitTests/Common/containers/) or [test cases](https://github.com/su2code/TestCases/tree/master/flamelet/01_laminar_premixed_ch4_flame_cfd).
 
 
-The artificial neural network type supported under the option `MLP` is the dense, feed-forward multi-layer perceptron. For FGM simulations, one or multiple MLP's can be loaded, over which the required flamelet manifold variables should be distributed. This offers the benefit of being able to use different MLP architectures for different data. `SU2` uses the [MLPCpp module](https://github.com/EvertBunschoten/MLPCpp.git) to evaluate MLP's during simulations. After training MLP architectures on flamelet data, it is possible to load these into `SU2` by storing the architecture, weights, biases, and activation function information in the supported `.mlp` format. Information on how to translate networks trained through TensorFlow to the `.mlp` format, see the [MLPCpp repository](https://github.com/EvertBunschoten/MLPCpp.git). Examples of such MLP files can be found in the `SU2` [unit tests](https://github.com/su2code/SU2/tree/master/UnitTests/Common/toolboxes/multilayer_perceptron), [test cases](https://github.com/su2code/TestCases/tree/master/flamelet/07_laminar_premixed_h2_flame_cfd), and the [MLPCpp repository](https://github.com/EvertBunschoten/MLPCpp.git).
+The artificial neural network type supported under the option `MLP` is the dense, feed-forward multi-layer perceptron. For FGM simulations, one or multiple MLP's can be loaded, over which the required flamelet manifold variables should be distributed. This offers the benefit of being able to use different MLP architectures for different data. `SU2` uses the [MLPCpp module](https://github.com/EvertBunschoten/MLPCpp.git) to evaluate MLP's during simulations. After training MLP architectures on flamelet data, it is possible to load these into `SU2` by storing the architecture, weights, biases, and activation function information in the supported `.mlp` format. Information on how to translate networks trained through TensorFlow to the `.mlp` format, see the [MLPCpp repository](https://github.com/EvertBunschoten/MLPCpp.git). Examples of such MLP files can be found in the `SU2` [unit tests](https://github.com/su2code/SU2/tree/master/UnitTests/Common/toolboxes/multilayer_perceptron), [test cases](https://github.com/su2code/TestCases/tree/master/flamelet/07_laminar_premixed_h2_flame_cfd), and the [MLPCpp repository](https://github.com/EvertBunschoten/MLPCpp.git). The neural networks were trained using the [SU2 DataMiner](https://github.com/EvertBunschoten/SU2_DataMiner) code. This repository contains test cases and tutorials which explain the data generation and set-up needed to train multi-layer perceptrons for flamelet-generated manifold applications.
 
 
 The preferred choice of manifold depends on the application, available computational resources, and available know-how. Typically, using a table results in shorter query times, and therefore simulation times compared to using MLP's. On the other hand, storing three-dimensional look-up tables with sufficient resolution for accurate FGM simulations requires substial amounts of memory (in the order of giga-bytes per core). If memory usage is not a constraint, using a LUT manifold therefore results in superior computational performance compared to using MLP's. The LUT uses linear interpolation over each cell in 2D queries. For highly non-linear data, this may result in inaccuracies and subsequent numerical instabilities during the simulation process. Using the MLP option allows for the use of non-linear activation functions, resulting in more smooth output data trends. The latter may result in improved solver robustness if the corresponding MLP can be trained to be sufficiently accurate. Finally, training artificial neural networks on flamelet data can be a cumbersome process and choosing an appropriate network architecture may not always be straightforward. On the other hand, generating a look-up table is comparatively easier.