Recall what "model" meant and lessons learnt:

Amplitude-versus-Offset (AVO)
Intercept and slope (gradient) multiplied by sine square of angle in a linearised model
Product of intercept and slope is an attribute (e.g., feature engineering)
Intercept and slope are two distinct features
Inversion
Gulf coast - low-frequency trend / young clastic sediments / cyclic deposition
North Sea - discrete-structure layers / carbonates / unconformities
Prestack Depth Migration
GOM - velocity gradient hanging off water bottom / varying with the depth of burial
North Sea - layer-cake model / distinct unconformities / Zechstein salt
Algorithms
Parameters optimization is an interpretive and iterative process.
Different model representations lead to very different workflow and algorithms.

Now consider machine-learning model more like linearised version of AVO, i.e., sum of weighted combination of features. Applying first principle based on sound geoscience will give us head start. Say we train machine to learn in the Middle East on the biggest reservoirs, algorithm may not do well in the deepwater Gulf of Mexico. Typical remedy will be to train on different data and eventually converge on a fit-for-purpose algorithm.

Geoscience in this case can save us time and focus on tuning the hyperparameters that create the possibility of breakthrough.