fix(docs): Sionna now uses Dr.Jit instead of TensorFlow.

Author: jeertmans

docs/source/motivations.md

@@ -28,7 +28,7 @@ DiffeRT builds on some of the principles behind DiffeRT2d while prioritizing per
 In terms of features, DiffeRT does not aim to match the extensive functionality of Sionna. Instead, DiffeRT focuses on RT-specific applications similar to what `sionna.rt` offers, but with four main differences:
 
 1. **Public lower-level RT Routines[^2]:** Many internal RT mechanisms in Sionna are hidden or undocumented, making it challenging to modify the pipeline. DiffeRT, on the other hand, ensures that most RT utilities are public and well-documented, enabling users to customize or replace parts of the RT algorithms without re-implementing or copy-pasting code.
-2. **JAX Integration:** Unlike Sionna, which uses TensorFlow, DiffeRT leverages JAX for efficient array-based programming. JAX offers powerful features like automatic differentiation, just-in-time (JIT) compilation, and compatibility with GPU/TPU acceleration, making it highly suitable for optimization and Machine Learning tasks.
+2. **JAX Integration:** Unlike Sionna, which uses Dr.Jit, DiffeRT leverages JAX for efficient array-based programming. JAX offers powerful features like automatic differentiation, just-in-time (JIT) compilation, and compatibility with GPU/TPU acceleration, making it highly suitable for optimization and Machine Learning tasks.
 3. **Minimal Abstraction with Immutable Dataclasses:** Sionna internally represents scenes using Mitsuba, which, while powerful, imposes restrictions on the types of scenes it can handle. Moreover, Sionna's classes are relatively complex, with many hidden attributes. In contrast, DiffeRT uses immutable dataclasses that can be created using simple constructors or convenient class methods (e.g., for reading scenes from files). Following JAX principles, all classes are immutable PyTrees, ensuring compatibility with JAX while avoiding unnecessary memory allocations through JIT optimization.
 4. **Lightweight and Broadcastable Design:** DiffeRT's design philosophy prioritizes transparency and usability for RT applications, avoiding the heavier abstractions often seen in other libraries. Classes aim to store as few attributes as possible, and most utilities accept input arrays with arbitrary sized inputs, which makes it very easy, e.g., to compute the same operation for one receiving (RX) antenna, or on a two-dimensional grid of RXs.