Three-Dimensional Surface Plots

Use a three-dimensional surface when height helps explain a function of two inputs. Align coordinate and height grids, control the projected viewpoint, compare surfaces fairly, and prefer contours when levels or paths need precise reading.

A three-dimensional surface plot uses two coordinate axes for a function's inputs and height for its output. For a loss with parameters weight and bias, one surface point has coordinates

(weight, bias, loss)

Height can make valleys, peaks, flat regions, and overall curvature intuitive. The surface is still projected onto a flat screen, however. Viewing angle, perspective, hidden regions, and axis scaling can make exact comparisons harder than they are in a contour plot.

Build one height for each coordinate pair

Use the same three grids introduced for contours:

All three arrays have shape (81, 81). At one grid position:

weight_grid[row, column] -> first horizontal coordinate
bias_grid[row, column]   -> second horizontal coordinate
loss_grid[row, column]   -> height

The shape agreement is not merely a Matplotlib requirement. It preserves the meaning that one weight and one bias produced one loss.

Draw a surface with height and color

Matplotlib creates a three-dimensional plotting area by passing projection="3d":

plot_surface then receives the three matching grids. Run the complete example and edit the coefficient of the bias term to change the bowl's shape.

The surface is a bowl with its lowest point at (2, -1, 0). Color repeats the height information and helps distinguish nearby regions. It does not add a fourth variable in this example.

Understand the flat projection

The figure appears three-dimensional because Matplotlib projects 3D coordinates onto the two-dimensional page. Some points lie behind others, and the visible size of a region depends on the viewpoint.

Set a reproducible view with:

axes.view_init(elev=30, azim=45)

elev is the elevation angle above the horizontal parameter plane. azim is the direction around the vertical axis. Changing either value changes the view, not the parameter grids or losses.

A low elevation can emphasize height but hide the far side of a valley. A high elevation approaches a top-down view but reduces the visible sense of height. Inspect more than one view while exploring, then choose and record the view that communicates the intended structure without hiding important regions.

Compare a surface with its contours

The same sampled function can be shown in both forms. The next example uses one set of coordinate and loss grids, so the difference comes from representation rather than data.

The surface makes the bowl-like geometry immediate. The contour plot makes equal values, parameter-plane distances, and possible optimization paths easier to compare. Use both when each answers a distinct question; otherwise prefer the simpler view.

Choose between a filled surface and a wireframe

plot_surface draws filled faces. A wireframe draws selected grid lines:

rstride and cstride select the step between displayed row and column grid lines. A wireframe can reveal sampling structure and leave more of the far side visible, but dense lines can become cluttered. A filled surface usually gives a clearer overall form. Choose from the inspection question, not from which style looks more elaborate.

Separate grid resolution from visual smoothness

A surface joins sampled grid values into faces. More coordinate samples can make those faces appear smoother:

21 × 21 grid -> 441 sampled points
81 × 81 grid -> 6,561 sampled points

The smoother rendering is not evidence that the underlying function or data is more accurate. Resolution should be fine enough to reveal the relevant structure without wasting computation. Sharp changes between samples can still be missed.

Surface plotting functions may also reduce the number of displayed samples for rendering efficiency. Keep the numerical grid and its resolution available when exact sampling matters.

Keep comparisons numerically fair

When comparing two surfaces, use the same:

• parameter ranges;
• coordinate sampling;
• horizontal and vertical axis limits;
• viewing angle;
• height limits;
• colormap and color limits.

Different height limits or viewpoints can make similar surfaces appear different. The apparent steepness also depends on the visual length assigned to each axis. Read slopes from numerical coordinates and loss changes rather than screen angle alone.

State what the surface represents

This example is genuinely a function of two parameters. A large model has many more. Its displayed surface may be:

• a toy loss designed for explanation;
• two selected parameters with all others fixed;
• or two selected directions through the full parameter space.

Different slices through the same high-dimensional loss can have different shapes. A smooth bowl in one slice does not prove that every direction is smooth, that the full problem has one minimum, or that optimization is easy.

Use height only when it improves the explanation

A surface plot is useful when visible height clarifies the overall geometry of a function of two inputs. Verify the three grid shapes, record the viewpoint and scales, compare important values numerically, and use contours when levels, paths, or parameter-plane distances need more precise reading.