A thin free plate obeys D∇⁴w + ρh ∂²w/∂t² = 0, D = Eh³/12(1−ν²). With the edges free (zero bending moment and zero effective shear) there is no closed-form mode shape for a square plate, so the modes here come from a Rayleigh–Ritz calculation: the deflection is expanded in products of free–free beam functions φi(x)φj(y) (8×8 terms) and the strain energy ½D&iint;[wxx² + wyy² + 2νwxxwyy + 2(1−ν)wxy²] and kinetic energy give a generalized eigenproblem Ka = Ω²Ma solved here in your browser (Cholesky + Jacobi). The free edge is a natural boundary condition of that energy, so it is satisfied automatically; ν = 0.3.
The frequency is f = Ω/(2πL²) · √(D/ρh). The nondimensional Ω is Leissa's λ²: 13.47, 19.60, 24.27… for the first elastic modes. The drive is a point force at the exciter: each mode responds as an = Fn/(Ωn²−Ω²+2iζΩnΩ) with ζ = 1/2Q and Fn ∝ wn(xd,yd) — so a mode with a node under the exciter is not excited (try the centre drive point), and between resonances the pattern is a blend of neighbouring modes. Sand sees the local amplitude |∑anwn|; it hops where Γ|w| > 1 (Γ = ω²w/g) and rests in the still band, so figures appear only near resonances.
Simplifications: 8×8 basis (the first dozen modes are accurate to a few tenths of a percent, the 20th less so); constant damping ratio; the exciter force is scaled so every resonance peaks at the same amplitude; sand hops are a random-walk stand-in for real bouncing and the instantaneous plate motion is shown in slow motion (1.2 Hz) with exaggerated displacement. This is different from the cosine-product figures of the sound/Chladni page, which ignore the free-edge condition. Bench, plate, exciter and vibrometer were modelled and baked in Blender.