diff --git a/moptipyapps/dynamic_control/objective.py b/moptipyapps/dynamic_control/objective.py
index 61a3a2f3..5866815c 100644
--- a/moptipyapps/dynamic_control/objective.py
+++ b/moptipyapps/dynamic_control/objective.py
@@ -8,7 +8,7 @@
 
 We offer two different approaches for this:
 
-- :class:`FigureOfMerit` computes the arithmetic mean `z `  over the separate
+- :class:`FigureOfMerit` computes the arithmetic mean `z` over the separate
   figures of merit of the training cases.
 - :class:`FigureOfMeritLE` tries to smooth out the impact of bad starting
   states by computing `exp(mean[log(z + 1)]) - 1`.
diff --git a/moptipyapps/dynamic_control/ode.py b/moptipyapps/dynamic_control/ode.py
index 0f74f6a1..b03ab590 100644
--- a/moptipyapps/dynamic_control/ode.py
+++ b/moptipyapps/dynamic_control/ode.py
@@ -550,7 +550,28 @@ def diff_from_ode(ode: np.ndarray, state_dim: int) \
     works reasonably well, then we could essentially plug this model into
     :func:`run_ode` instead of the original `equations` parameter.
 
-    :param ode: the result of :func:`ode_run`
+    What this function does to compute the differential is to basically
+    "invert" the dynamic weighting done by :func:`run_ode`. :func:`run_ode`
+    starts in a given starting state `s`. It then computes the control vector
+    `c` as a function of `s`, i.e., `c(s)`. Then, the equations of the dynamic
+    system (see module :mod:`~moptipyapps.dynamic_control.system`) to compute
+    the state differential `D=ds/dt` as a function of `c(s)` and `s`, i.e., as
+    something like `D(s, c(s))`. The next step would be to update the state,
+    i.e., to set `s=s+D(s, c(s))`. Unfortunately, this can make `s` go to
+    infinity. So :func:`run_ode` will compute a dynamic weight `w` and do
+    `s=s+w*D(s, c(s))`, where `w` is chosen such that the state vector `s`
+    does not grow unboundedly. While `s` and `c(s)` and `w` are stored in one
+    row of the result matrix of :func:`run_ode`, `s+w*D(s,c(s))` is stored as
+    state `s` in the next row. So what this function here basically does is to
+    subtract the old state from the next state and divide the result by `w` to
+    get `D(s, c(s))`. `s` and `c(s)` are already available directly in the ODE
+    result and `w` is not needed anymore.
+
+    We then get the rows `s, c(s)` and `D(s, c(s))` in the first and second
+    result matrix, respectively. This can then be used to train a system model
+    as proposed in model :mod:`~moptipyapps.dynamic_control.system_model`.
+
+    :param ode: the result of :func:`run_ode`
     :param state_dim: the state dimensions
     :returns: a tuple of the state+control vectors and the resulting
         state differential vectors
diff --git a/moptipyapps/dynamic_control/system_model.py b/moptipyapps/dynamic_control/system_model.py
index 01b4bd78..d2307469 100644
--- a/moptipyapps/dynamic_control/system_model.py
+++ b/moptipyapps/dynamic_control/system_model.py
@@ -6,8 +6,11 @@
 controller output `control` for the state vector `state`. It will return
 the differential of the system state, i.e., `dstate/dT`. In other words,
 the constructed model can replace the `equations` parameter in
-:func:`~moptipyapps.dynamic_control.ode.run_ode`. The idea is here to
-re-use the same function models as used in controllers
+:func:`~moptipyapps.dynamic_control.ode.run_ode`. The input used for
+training is provided by
+:func:`~moptipyapps.dynamic_control.ode.diff_from_ode`.
+
+The idea here is to re-use the same function models as used in controllers
 (:mod:`~moptipyapps.dynamic_control.controller`), learn their
 parameterizations from the observed data, and wrap everything together
 into a callable.