Advanced · 15 min

The periodic table as a data landscape

Render periodic properties — atomic radius, electronegativity, ionisation energy — as 3D terrain and practise spotting the periodic trends a machine-learning model would have to learn.

Goal

Turn the periodic table into a regression landscape: identify features a model could use (period, group) and the irregularities it would struggle with.

Apparatus and reagents

A grid of elements lifted into 3D, with surface height proportional to the selected periodic property.

Procedure

Keep mode 0 and rotate the landscape; find the valley or ridge that marks each period and group.
Switch to mode 1 and compare: does the property rise or fall along a period, and is the trend smooth or jagged?
Switch to mode 2 and look for outliers — elements that break the monotonic trend.
For each mode, state the rule a model should learn (e.g. “increases up and to the right”) and name one exception.

What to observe

Periodic properties vary smoothly along rows and columns — neighbours in the table are similar, exactly what descriptor-based ML exploits.
Some modes show sharp ridges or drops at the transition metals and lanthanides — regions where simple linear models would fail.
The same element occupies the same (row, column) in every mode — a reminder that structure, not chemistry alone, carries predictive information.

Explanation

Machine learning in chemistry often starts with descriptors: numerical features such as atomic number, group, period or electronegativity that a model regresses against a target property. Viewed as a landscape over the periodic table, a learnable property is a smooth surface; a hard property is rugged. This is why featurisation matters and why graph or kernel models that know the periodic structure outperform naive regressions on Z alone.

Chemists behind it

Dmitri Mendeleev John Pople