diff --git a/README.md b/README.md
index c7d072a..f37cd03 100644
--- a/README.md
+++ b/README.md
@@ -105,8 +105,7 @@ python run_pipeline.py --cfg_file configs/WBC-SPH.yml \
 The data was generated based on the code of previous work (https://github.com/isl-org/DeepLagrangianFluids). The required code is in the 'datasets' folder. 'create_data.sh' is the shell script to run the data generation.
 
 ### Why not replace all the CConv with the ASCC?
-The reason for this is that the antisymmetry is a severe restriction. This makes the problem much more difficult for the neural network to solve. It is sufficient to place the ASCC only at the end to make the network antisymmetric. Even with this, it was very difficult to adjust the mesh to the current state. On the other hand, the network can generalized much better. The antisymmetric layer can also be seen as a kind of mapping into a reduced antisymmetric space. 
-The paper has a short paragraph with an example (Standing Liquid) in the Result section, which briefly discusses this.
+The reason for this is that the antisymmetry is a strong restriction. Replacing all CConvs with ASCCs severely limits the learning capabilities of the neural network. The good thing is, it is sufficient to place the ASCC only at the end to make the network antisymmetric. This reduces the constraints in the network to a minimum and only transforms the generated values in the final layer so that the resulting values comply with the antisymmetry constraint. Even with this, it was quite difficult to tune the network to achieve the current state. The paper has a short paragraph with an example (Standing Liquid) in the Result section, which briefly discusses this.
 
 ### What is the *Maximum Density* in the evaluation?
 The *Maximum Density* value is the relative error between the maximum density of the fluid and the maximum density of the ground truth, where a value closer to 0 is preferable (Equation 15 in the paper). We use this as a heuristic for the compressibility of the fluid, which can lead to high pressure and thus instability in the simulation. Apart from that, please note that the values in Table 2 in the paper are not the raw error values but relative accuracy values as described in Figure 6. I.e. a value of 1 corresponds to the error of our final method, while small values represent a lower relative accuracy and thus larger error. A value of 0.5, for example, would mean half the accuracy and double the error. We chose this format to relate the error to the final method, which we felt was important in an ablation study, and to normalise the error evaluation for better visualisation in the graph.