CHAPTER 3

Discrete Element Method Theory

Equations of motion, contact geometry, force models and numerical stability criteria used by the BrixDEM solver.

In the Discrete Element Method, each particle is tracked individually and its motion is obtained by explicit integration of Newton's equations. Forces arise mainly from contacts: when two spheres overlap numerically, this overlap is interpreted as local elastic deformation and produces normal and tangential forces. The quality of the simulation depends directly on the choice of material parameters, the contact model and the timestep Δt.

The soft-sphere formulation adopted by BrixDEM follows the classical DEM literature, initiated by Cundall & Strack [1] and consolidated in the contact models of Hertz [2] and Mindlin [3, 4]. Each equation below includes the corresponding primary reference from the scientific literature; the complete list is in Appendix F.

3.1 Equations of motion

For each particle i of mass mi, radius Ri and moment of inertia Ii, BrixDEM solves linear and angular momentum balances over all active contacts. Integration uses explicit velocity-Verlet integration [18], with normal, relative velocity and tangential history evaluated consistently in particle-particle contact:

equation 3.1(3.1)
equation 3.2(3.2)
equation 3.3(3.3)

LITERATURE[1] Cundall & Strack (1979) - soft-sphere DEM formulation; [18] Verlet (1967) - explicit integration. Solid sphere moment of inertia I = 2/5 mR2.

Contact forces and torques come from the selected model (Hertz–Mindlin or Linear Spring–Dashpot), summed over all active particle contacts.

3.2 Contact geometry

For two spherical particles i and j, there is contact when the distance between centers is less than the sum of their radii. The vector direction between centers defines the contact normal, which separates the problem into a normal component (compression/restitution) and a tangential (friction/slip).

equation 3.4(3.4)
equation 3.5(3.5)
equation 3.6(3.6)
equation 3.7(3.7)

LITERATURE[1] Cundall & Strack (1979); [19] Schäfer, Dippel & Wolf (1996) - force schemes and contact kinematics in granular media.

Figure 3.1 - Pairwise contact model. The δn overlap between the two spheres generates the normal force Fn (kn spring and cn dashpot along normal n) and the tangential force Ft (kt spring and dashpot, ct limited by Coulomb friction). It is the classic representation of contact in DEM books.
Figure 3.1 - Pairwise contact model. The δn overlap between the two spheres generates the normal force Fn (kn spring and cn dashpot along normal n) and the tangential force Ft (kt spring and dashpot, ct limited by Coulomb friction). It is the classic representation of contact in DEM books.

Relative speed includes translation and rotation, correctly capturing shear on the contact surface. The small overlap δn does not represent physical interpenetration - it is the idealized local elastic deformation.

Physical interpretation of contact elements

ELEMENTS PHYSICAL MEANING ASSOCIATED PARAMETER
Overlap δₙ Virtual elastic deformation; generates normal repulsion force Normal stiffness kn
Normal relative velocity vn Approach or separation of particles Restitution eₙ, damping cₙ
Tangential velocity vt Relative sliding at the contact surface Friction μ, stiffness kt, damping ct
Angular velocity ω Particle rotation during contact Rolling μ_r, torsion μ_t
Normal force Fn Compression/repulsion force between bodies kn, cn, E, ν, δn
Tangential force F t Resists sliding up to the Coulomb limit μ, k t, c t

3.3 Restitution coefficient

The normal restitution coefficient eₙ measures how much of the normal speed is recovered after impact. The solver uses β ≥ 0 and calculates dissipative coefficients from en; therefore, lower en values produce greater damping:

equation 3.8(3.8)
equation 3.9(3.9)

LITERATURE[5] Tsuji, Tanaka & Ishida (1992) - restitution-dependent damping; [6] Di Renzo & Di Maio (2004) - comparison of force models and ratio β(en).

In a simple validation by free fall - a particle falling from height h₁ and rebounding to h₂ in a purely normal collision — the following approximate relation holds:

equation 3.10(3.10)
Figure 3.2 - Restitution coefficient. The particle strikes with a normal incoming velocity and rebounds at a lower outgoing velocity; the ratio of fall height h₁ and rebound height h₂ gives eₙ ≈ √(h₂/h₁).
Figure 3.2 - Restitution coefficient. The particle strikes with a normal incoming velocity and rebounds at a lower outgoing velocity; the ratio of fall height h₁ and rebound height h₂ gives eₙ ≈ √(h₂/h₁).

en = 1 represents ideally elastic collision; en = 0, ideally plastic/dissipative. This relationship is useful as a sanity test and is the basis of the rebound calibration assistant (Chapter 18).

3.4 Tangential Friction and Coulomb Limit

The friction configured in the interface acts as contact sliding friction. The tangential model calculates a trial force F_t* that resists the relative slip. As long as this force remains below the Coulomb limit, contact remains in stick; beyond the limit, there is slip.

equation 3.11(3.11)
equation 3.12(3.12)

LITERATURE[3] Mindlin (1949); [4] Mindlin & Deresiewicz (1953) - tangential stiffness and contact friction; [1] Cundall & Strack (1979) - Coulomb friction limit in DEM.

Physically, the μ coefficient controls the tendency of the material to slip or adhere. It governs the formation of piles, the angle of repose, shear in chutes and hoppers, and stability in silos and mixers.

How the solver applies these equations — Three implementation details worth knowing: (1) the trial force tested against the Coulomb limit includes the tangential dashpot term (eq. 3.11), not just the spring — and, when sliding, the history Δs_t is rescaled consistently with the saturated force; (2) the limit μ‖Fₙ‖ uses the pair's effective normal force, which includes the compressive cohesion contribution when active (Chapter 3.9) — cohesive material has greater frictional resistance; (3) the normal contact never "pulls": if, in the final phase of the contact, the dashpot would make Fₙ attractive, the solver zeroes it — normal tension only exists in the cohesion models.

3.5 Effective properties of the pair of materials

When two bodies interact (particle-particle or particle-wall), the solver uses effective properties calculated from individual properties:

equation 3.13(3.13)
equation 3.14(3.14)
equation 3.15(3.15)

LITERATURE[2] Hertz (1882) - effective modulus E of elastic contact; [3] Mindlin (1949) - effective shear modulus G.

In BrixDEM the contact parameters of the pair - restitution coefficient en, friction μ, rolling friction μ r and torsion μ t - are directly defined for each interaction in the Interactions tab (Chapter 10), and not obtained from mixing rules between materials. The effective geometric and elastic properties above (E, G, R, m) are the natural physical combinations of the pair and are calculated by solver from the properties of each body.

3.6 Hertz–Mindlin model

Hertz–Mindlin is a nonlinear model: effective stiffness increases with the square root of overlap δₙ. It is recommended when you want greater physical fidelity in elastic contacts - particularly where the duration of contact and the transmission of force affect the result.

equation 3.16(3.16)
equation 3.17(3.17)
equation 3.18(3.18)
equation 3.19(3.19)

LITERATURE[2] Hertz (1882) — nonlinear normal stiffness kₙ ∝ √(R*δ); [3] Mindlin (1949) — tangential stiffness; [5] Tsuji et al. (1992) and [6] Di Renzo & Di Maio (2004) — damping based on restitution coefficient eₙ.

The tangential component uses the history of tangential displacement Δs_t. In particle-particle contact, this history is carried by an objective rotation between the previous and current normals, which preserves its norm when the contact plane rotates and prevents spurious dissipation caused by simple projection:

equation 3.20(3.20)
equation 3.21(3.21)

LITERATURE[4] Mindlin & Deresiewicz (1953) - incremental tangential loading with displacement history; [17] Thornton (2015) - objective formulation of contact kinematics.

3.7 Linear Spring–Dashpot model

The Linear Spring–Dashpot (LSD) model uses constant stiffness. It is widely used in calibrations, large industrial simulations and studies where the user wants to directly control stiffness and collision time:

equation 3.22(3.22)
equation 3.23(3.23)

LITERATURE[1] Cundall & Strack (1979) - original linear spring–dashpot model; [19] Schäfer, Dippel & Wolf (1996) - linear force schemes.

Figure 3.3 — Loading (blue) and unloading (orange) cycle of the normal force in a damped collision, for the linear model (left) and Hertz–Mindlin (right). The dotted line is the purely elastic response; the loop area is the energy dissipated by the dashpot. Note that unloading ends at Fₙ = 0 before δₙ returns to zero — the solver does not allow an attractive normal force (tension only via the cohesion models).
Figure 3.3 — Loading (blue) and unloading (orange) cycle of the normal force in a damped collision, for the linear model (left) and Hertz–Mindlin (right). The dotted line is the purely elastic response; the loop area is the energy dissipated by the dashpot. Note that unloading ends at Fₙ = 0 before δₙ returns to zero — the solver does not allow an attractive normal force (tension only via the cohesion models).

In automatic mode, normal stiffness can be estimated from a Reference overlap user-defined δʳᵉᶠ, calibrating the linear model to reproduce the Hertz–Mindlin contact time. The tangential/normal ratio kt = 2/7 kn is the classic ratio for elastic spheres:

equation 3.24(3.24)
equation 3.25(3.25)
equation 3.26(3.26)

LITERATURE[6] Di Renzo & Di Maio (2004) - equivalent linearisation of Hertz; [16] O'Sullivan & Bray (2004) - choice of stiffness/step; [21] Walton & Braun (1986) - ratio kt = 2/7 kn.

A point in favour of LSD for calibration: in the linear oscillator, the relation between the damping cₙ and the restitution coefficient eₙ (equations 3.9 and 3.26) is exact — the analytic solution of the damped collision — whereas in Hertz–Mindlin the same relation is an approximation. With LSD, the eₙ you configure is exactly the eₙ the normal collision reproduces.

Sensitivity to δʳᵉᶠ — In automatic mode, kₙ grows with √δʳᵉᶠ — the stiffness (and with it the collision time and the admissible Δt) depends on the chosen reference overlap. Use values around ~0.1–1 % of the effective radius and, after running, check in the charts that the actual overlap (Overlap/R) stayed close to the reference; if it came out much higher, increase kₙ (or reduce δʳᵉᶠ) and repeat.

In advanced mode, kn, cn, k t and c t can be entered manually. When defined, manual values prevail over automatic estimates.

When to use each model - Use Hertz–Mindlin when the case is small, with fast and elastic contacts (impacts, fall of individual particles). Use LSD in large industrial cases or for calibration, because it allows direct control of the collision time and admits larger Δt.

3.8 Rolling and torsion resistance

Perfect spheres roll with very little resistance - which produces unrealistic angles of repose and piles that “flow”. To reproduce real angular grains, DEM adds two resistant torques: a rolling resistance (rolling), which opposes relative rolling in the contact plane, and the torsion resistance (torsion), which opposes the spin around normal. BrixDEM implements the three rolling models of the Ai et al. (2011) taxonomy — Model A (CDT), Model B (viscous) and Model C (EPSD) — and one physical torsion model, Marshall (2009), all selectable per pair of materials in the Interactions tab (see Chapter 10).

Common kinematics

Both torques derive from the relative angular velocity of contact, decomposed into a component of rolling (in the contact plane) and a component of spin (along the normal):

equation 3.27(3.27)
equation 3.29(3.29)

LITERATURE[11] Ai, Chen, Rotter & Ooi (2011) — taxonomy and assessment of rolling resistance models; [12] Iwashita & Oda (1998) — rolling kinematics at the contact.

Figure 3.4 - Kinematics of rolling resistance and torsion. Rolling acts in the contact plane (ω^roll), opposing the relative spin of the spheres; torsion acts about the contact normal (ω^spin).
Figure 3.4 - Kinematics of rolling resistance and torsion. Rolling acts in the contact plane (ω^roll), opposing the relative spin of the spheres; torsion acts about the contact normal (ω^spin).

Rolling models

CDT — Constant Directional Torque (Model A of Ai et al.). Applies a resisting torque of constant magnitude, proportional to μ_r, R* and the normal force, always opposing the rolling angular velocity. It is the standard model — robust and cheap:

equation 3.28(3.28)

LITERATURE[11] Ai et al. (2011) — Model A (Eq. 8a); [20] Zhou et al. (1999) — rolling friction in sandpile formation.

The characteristic limitation of CDT shows up at rest: since the torque has fixed magnitude and only flips sign with the rolling direction, a particle that should stop keeps oscillating around its final position at a frequency set by the time step — torque and friction force flicker between ±M. In dynamic flow this is irrelevant; in quasi-static beds it can contaminate force averages and prevent complete rest (exactly the defect Model C fixes).

EPSD — Elastic-Plastic Spring-Dashpot (Model C of Ai et al., Eqs. 10–20). Accumulates the relative rolling angle θ_r in a torsional spring and applies a torque proportional to that history. While the torque stays below the plastic limit it is reversible (rolling adherence); on reaching the limit μ_r R* ‖Fₙ‖ it saturates by return-mapping (plastic yielding), exactly like tangential friction:

equation 3.28a(3.28a)
equation 3.28b(3.28b)

The rolling stiffness is computed per contact from the normal force — it is not a global number to calibrate. BrixDEM uses the theoretical contact stiffness of Ai et al. (Eq. 7c, after Bardet & Huang) with J_n = ½, which is equivalent to saying that the torque fully mobilises after rolling through a mobilisation angle θ_m = μ_r:

equation 3.28d(3.28d)

Below full mobilisation a critical dashpot with damping ratio ζ (0–1) also acts, dissipating the spring oscillation and allowing complete rest; at full mobilisation the dashpot is switched off (f = 0 in Eq. 17 of Ai et al.). The moment of inertia I_r is that of the rolling vibration mode about the contact point (Ai, Eq. 20) — for a solid sphere each body contributes 7/5 m R²:

equation 3.28e(3.28e)

LITERATURE[11] Ai et al. (2011) — Model C (Eqs. 10–20): elastic-plastic spring with a return-mapping curve, per-contact stiffness and critical dashpot; [12] Iwashita & Oda (1998) — origin of the elastic-plastic rolling model; [13] Wensrich & Katterfeld (2012) — rolling friction as a shape proxy.

Behaviour verified in the solver’s analytic suite: a sphere launched in pure rolling decelerates at a = (5/7)μ_r g, travels d = (7/10)v₀²/(μ_r g) and comes to rest with no residual oscillation; on a ramp inclined less than the resistance angle β = arctan(μ_r) the sphere stays put (Ai, Eq. 1). It is the recommended model for angle of repose, packing and quasi-static regimes. The field k_r > 0 remains available as a manual stiffness override to reproduce older studies; leave k_r = 0 for the automatic stiffness (3.28d).

Viscous (Model B of Ai et al., Eq. 9; proposed by Zhou et al.). Applies a torque proportional to the rolling velocity at the contact surface — the term (R_i ω_i − R_j ω_j), projected onto the contact plane — scaled by the normal force:

equation 3.28c(3.28c)

LITERATURE[11] Ai et al. (2011) — Model B (Eq. 9); [20] Zhou et al. (1999) — original viscous model.

Note the sign combination: for two spheres in mutual rolling (opposite ω) the contributions add and the torque resists rolling; for gear-like spinning (same ω — the surfaces slide) the term vanishes and dissipation is correctly left to tangential friction. For a sphere rolling on a plane the model yields exponential decay, v(t) = v₀·e^(−(5/7)μ_r g t) — with no static threshold. Model B therefore cannot sustain piles nor produce a realistic angle of repose (a conclusion of Ai et al.’s own paper); use it to damp rotations in dynamic regimes, not in quasi-static problems.

Torsion model — Marshall

Torsion resistance uses the formulation of Marshall (2009, Eqs. 29–34): a spring and a dashpot on the relative spin angle θ^spin whose coefficients are not user parameters — they are derived from the tangential contact itself, by integrating the friction traction over the circular contact patch of radius a. The torsional stiffness k_Q and damping η_Q inherit the tangential stiffness k_t and damping c_t of the force model, scaled by a²/2, and the torque saturates at a torsional Coulomb limit with lever arm (2/3)a:

equation 3.30(3.30)
equation 3.31(3.31)
equation 3.31b(3.31b)

LITERATURE[14] Marshall (2009) — Eqs. 29–34: k_Q = k_t a²/2, η_Q = c_t a²/2 and M_crit = (2/3)·a·μ‖Fₙ‖ by integrating the friction traction over the patch; [3] Mindlin (1949) — tangential stiffness k_t = 8G*a from which k_Q derives.

The coefficient μ in the limit is the pair’s sliding friction itself: in the paper, F_crit = μ‖Fₙ‖ is the Coulomb critical force of sliding (Eq. 25) and the critical torque results from integrating that traction over the patch (Eq. 33). That is why the interface does not ask for a “torsion μ” — when torsion is enabled, the solver automatically uses the interaction’s μ. Above the limit the contact enters torsional sliding (M = −M_crit along the spin direction, with the same return-mapping as tangential friction); when sliding ceases, the spring k_Q captures the spin and the dashpot η_Q dissipates the residual oscillation. Analytic check: a sphere spinning on a plane decelerates at α = M_crit/I = (5/3)·a·μ·g/R² while sliding, then comes to rest.

Limitations — force-model dependence — Torsion inherits k_t, c_t and the patch radius a from the normal/tangential contact model: with Hertz–Mindlin, k_t = 8G*a and a = √(R*δₙ) — Marshall’s formulation holds in full; with JKR, the solver uses the actual adhesive contact radius a(δₙ), consistent with the theory. With LSD, however, k_t and c_t are user constants and the patch uses only the geometric radius a = √(R*δₙ): k_Q and η_Q lose the Mindlin scaling and the model becomes a phenomenological approximation — the qualitative behaviour (spring + saturation + damping) is preserved, but the torque magnitude should be treated as a calibration quantity. Under any force model, torsion resistance vanishes as the contact unloads (a → 0) and the maximum torque (2/3)aμ‖Fₙ‖ is orders of magnitude smaller than the rolling one (lever a ≪ R*): it is a fine effect — relevant for dense packings, interlocked grains and torsional shear; secondary in dilute flow.

Which model to choose

MODEL TYPE PARAMETERS WHEN TO USE
Rolling · CDT (Model A) Friction (Coulomb) μ_r Default for dynamic flow: piles, drums; oscillates at rest
Rolling · EPSD (Model C) Elastic-plastic spring + dashpot μ_r, ζ (k_r automatic) Angle of repose, static packing, quasi-static; rests cleanly
Rolling · Viscous (Model B) Torque ∝ rolling velocity μ_r Damp rotations in dynamic regimes; cannot sustain piles
Torsion · Marshall Spring + dashpot derived from the contact — (uses the pair’s μ) Dense packing, interlocked grains, torsional shear
Calibration — μ_r has a large effect on the angle of repose. For spherical grains, use a non-zero μ_r (CDT is a good starting point in flow; EPSD when the material must truly come to rest) and validate with the pile test (Calibration › Angle of repose assistant). In EPSD, leave k_r = 0 — the stiffness is automatic per contact (eq. 3.28d) — and set the damping ratio ζ between 0 and 1: 0.3 is the reference value of Ai et al. (2011); with ζ = 0 the dashpot is off and the sphere oscillates indefinitely at its final position. Marshall torsion has no parameters of its own: enable it when spin about the normal matters; it uses the pair’s μ.

3.9 Cohesion and adhesion

Cohesion adds an attractive normal contribution to the contact, representing van der Waals forces, liquid bridges or surface adhesion. It increases the stability of agglomerates, profoundly changes the flow and, for very sticky materials, requires a smaller Δt. In the additive formulation, the cohesive force is added to the normal contact force:

equation 3.32(3.32)

LITERATURE[31] Coetzee & Scheffler (2023) — review of DEM cohesion models and their calibration for cohesive materials; [7] Johnson, Kendall & Roberts (1971) — adhesion of elastic solids.

BrixDEM implements two models, chosen by the compatibility with the normal force model:

MODEL COMPATIBLE WITH NATURE HYSTERESIS
LinearCohesion (linear additive) LSD and Hertz–Mindlin Additive force F = k·A, in overlap No
JKR (complete) Hertz–Mindlin; P–P and P–W (not W–W) Replaces Hertz normal; full adhesion + pull-off, in P–P and in P–W Yes

LinearCohesion — linear additive cohesion

Adds an attractive force proportional to the contact area A — the intersection circle of the two spheres — acting only while there is overlap (δn > 0). It is a simple additive model: cheap, robust and valid for both the linear (LSD) model and Hertz–Mindlin. The parameter k is the cohesion energy density [J/m³] — the larger the compressed area, the stronger the attraction, so cohesion grows smoothly with overlap and vanishes on separation.

equation 3.39(3.39)

For particle-wall contact, the area uses the sphere–plane formula and the attraction points to the wall:

equation 3.40(3.40)

LITERATURE[31] Coetzee & Scheffler (2023) — taxonomy of DEM cohesive models (the simplified SJKR F = k·A among them) and calibration from bulk response; [7] Johnson, Kendall & Roberts (1971) — adhesive basis.

Limitations of LinearCohesion — The attraction only exists in overlap (δₙ > 0) and goes to zero on separation: there is no finite pull-off force, no bond hysteresis and no traction across an open gap. The model cannot reproduce the abrupt “detachment” of fine powders nor agglomerates that resist tension — use JKR for those effects. Treat k [J/m³] as an effective parameter, calibrated against the material’s bulk response (angle of repose, flow rate, shear cell), as recommended by the cohesive-calibration literature [31] — not as a measured physical property.

JKR — Johnson–Kendall–Roberts (complete)

The JKR model replaces the Hertz normal force with an adhesive law parameterized by the contact radius a. Here, a is the radius of the contact circle — the flattened, approximately circular region that forms where the two spheres press against each other. In pure Hertz contact a = √(R*·δn); in JKR the adhesion deforms the edges of that circle (an adhesive “neck” forms) and a is larger than the Hertz value for the same overlap δn. The material parameter is γ, the surface energy [J/m²]: the energy gained per unit area when the two surfaces adhere. The theory follows from the balance between the Hertz elastic energy and the surface term (−γ·πa²), and this competition produces the JKR signatures: a finite contact radius at zero load (a > 0 at δn = 0), sustained tension across an open gap (δn < 0) and a pull-off force on separation. In BrixDEM it can be used on both particle–particle and particle–wall contacts (not wall–wall, which is not a resolved contact) — and both show the full hysteretic behavior: the bond forms at first touch, sustains tension across an already open gap and breaks at pull-off. The only difference between them is the effective radius: for two equal spheres R_eff = R/2, against a flat wall R_eff = R — so the pull-off force at the wall is double that of the particle–particle pair with the same γ (Chapter 28.4; verified in Appendix H, Case 8b). It is recommended for fine powders and cases where elastic adhesion dominates the dynamics. The law is given in parametric form — overlap and force written as functions of a:

equation 3.41(3.41)
equation 3.42(3.42)

LITERATURE[7] Johnson, Kendall & Roberts (1971) - original JKR theory (surface energy and adhesive contact).

The adhesive bond only breaks when the traction reaches the pull-off, at a negative distance δ pull (open gap). The inversion a(δₙ) uses the closed-form analytical solution of Parteli & Pöschel (2014):

equation 3.43(3.43)

LITERATURE[9] Parteli et al. (2014) - closed contact radius a(δ) solution for fine powders; [7] JKR (1971) - pull-off force 3πγR*.

Between different materials, surface energy is combined by geometric mean. Damping comes from the coefficient of restitution eₙ through the Tsuji formulation (β from Eq. 3.9), with normal stiffness evaluated at the actual JKR contact radius, Sₙ = 2E*a. Mindlin tangential stiffness uses the same radius; tangential damping uses the Tsuji form with kₜ:

equation 3.44(3.44)

LITERATURE[5] Tsuji, Tanaka & Ishida (1992) — restitution-dependent damping; [3] Mindlin (1949) — tangential stiffness using the contact radius.

Hysteresis and timestep - JKR maintains a binding state (bonded) per pair: adhesion is born at first touch (δn ≥ 0) and only breaks in the pull-off (δn < −δ pull). Cohesive models strongly alter flow, packing and stability. Perform a Δt convergence study and monitor kinetic energy, overlap and particle count. The "total energy" quantity of the standard output does not include all elastic/adhesive energy stored in contacts.

3.10 Timestep - Rayleigh criterion

The timestep Δt is one of the most sensitive parameters in DEM. A Δt too large generates excessive overlap, artificial energy, exploding particles and numerical instability. In BrixDEM the automatic calculation of Δt is based exclusively on Rayleigh time - the time for an elastic surface wave to travel through the particle. Solver calculates t R for each particle material and uses the smallest material value, not just the smallest Young's modulus:

equation 3.33(3.33)
equation 3.34(3.34)
equation 3.35(3.35)

LITERATURE[15] Thornton & Randall (1988) — critical time step from the Rayleigh surface wave, with C_R ≈ 0.1631ν + 0.8766.

Automatic timestep

The mode Rayleigh (automatic) applies the α R factor to the shortest Rayleigh time between active materials:

equation 3.36(3.36)

LITERATURE[15] Thornton & Randall (1988) — Rayleigh criterion; [16] O'Sullivan & Bray (2004) — time-step verification by convergence (their criterion is the mass–spring one, not Rayleigh's). The current automatic mode of BrixDEM uses Rayleigh only.

SCENARIO α R INITIAL RECOMMENDATION
Simple, monodisperse case 0,20 Confirm by Δt convergence
Polydisperse, small particles 0,10–0,15 Use the smallest radius effectively present
Very rigid material 0,10 Monitor overlap and kinetic energy
Active cohesion (JKR/LinearCohesion) ≤ 0,10 Reduce additionally and validate pull-off
PSD and minimum radius - Rayleigh's limit depends linearly on the radius. In a PSD, the safe value must use the smallest radius that can be generated. Confirm that the reference radius of the material and the minimum PSD are consistent; this condition is audited separately in the technical report.
Excessive Δt symptoms - Kinetic energy grows without physical cause, particles cross walls, overlaps greater than ~1% of the radius, messages of instability in the console. Halve Δt and test again before any other action.

3.11 Contact with walls, STL and primitives

Walls, STL meshes and geometric primitives use the same contact logic as particles: local normal, overlap, relative speed, normal force, tangential force and torques. The difference is that the wall is treated as a fixed or prescribed body, while the particle responds dynamically.

equation 3.37(3.37)
equation 3.38(3.38)

LITERATURE[1] Cundall & Strack (1979); [19] Schäfer et al. (1996) - Particle–boundary contact using the same force law; wall treated as a body of infinite mass/radius.

3.12 Coarse-graining: scaling laws

Coarse-graining uses one larger particle to represent a group of grains. Multiplying radius by k at constant density makes representative-particle mass grow as k³. At the same total mass, ideal particle count decreases by k³:

equation 3.45(3.45)
equation 3.46(3.46)

These are geometric relations; computing-time reduction also depends on contacts, timestep, output and spatial distribution. In the reference scale with unchanged E, density ρ [kg/m³] and modulus E [Pa] remain constant, as do dimensionless coefficients ν, e and μ:

equation 3.47(3.47)

For JKR, scaling surface energy γs by k makes pull-off grow as k² because effective radius also grows by k:

\gamma_{s,\mathrm{CG}}=k\gamma_s\quad\Longrightarrow\quad F_{\mathrm{pull,CG}}=k^2 F_{\mathrm{pull}}(3.48)

This does not preserve the individual adhesion-to-weight ratio: weight grows as k³ at fixed gravity. Preserving interaction stress, individual Bond number and a trajectory in fixed-size equipment are different similarity objectives. One scaling rule cannot impose all of them without checking its assumptions.

The PSD, including its bounds, is geometrically enlarged:

equation 3.49(3.49)

At constant ρ and ν, Rayleigh time is proportional to radius and inversely proportional to the square root of E:

\frac{t_{R,\mathrm{CG}}}{t_R}=k\sqrt{\frac{E}{E_{\mathrm{CG}}}}\qquad(\rho,\nu\ \mathrm{const.})(3.50)

Thus unchanged E gives critical time ×k; E divided by k² gives critical time ×k². This is a contact-stability relation, not proof of process-time equivalence. Optional E reduction softens contacts and requires deformation/calibration checks; its states and the distinction between entered and effective dt are described in 10.7.

LITERATURE[26] — Particle-system similarity laws under scaling assumptions. The reference does not universally justify optional E reduction or equivalence in fixed equipment and gravity.

Required comparison. Compare the CG case against CGF = 1 for relevant flow rate, velocity profile, segregation and power/torque. With breakage, preserve the work convention and calibrate by size/shape; with liquid, use the specific scaling in 3.15 and local checks in Appendix H. Agreement in one observable does not guarantee agreement in others.

3.13 Surface wear: physical models

BrixDEM estimates wear of surfaces (chutes, liners, paddles, screens) from the forces and slip of the particle-wall contacts, accumulating the depth removed [m] per mesh triangle. Two physical mechanisms can be combined.

Abrasive wear (Archard)

Models removal proportional to the friction work of sliding contact. Each step accumulates a depth proportional to normal force and sliding displacement, normalized by triangle area (A△). K/H is the specific wear coefficient [1/Pa]:

equation 3.51(3.51)

LITERATURE[22] Archard (1953) - volume removed proportional to load × sliding distance / hardness.

Impact erosion (Oka)

Models the removal caused by particles that collide with the surface, evaluated at the moment of impact. It depends on impact angle α (function g(α)) and speed v:

equation 3.52(3.52)
equation 3.53(3.53)

g(α) describes angular dependence; E90 is the reference erosion at 90°, function of Vickers H v hardness, velocity v and diameter D, with calibrable exponents. The depth is normalized by the area of the triangle and scaled by particle mass.

LITERATURE[23] Oka, Okamura & Yoshida (2005) - predictive equation of impact erosion (angle, speed, hardness); [25] Maw, Barber & Fawcett (1976) - oblique impact of elastic spheres.

How to use it in the GUI - Activation and wear parameters (including default values of the Oka exponents) and the interpretation of the results are in Chapter 29.
Where is the full literature? - Each equation in this chapter quotes its primary reference in square brackets. The full numbered list is in Appendix F · References.

3.14 Composite particles (clumps): fundamentals

A clump is a rigid body made of N spheres ("pebbles") of radii rᵢ, with controlled overlap between them, that translate and rotate as a single body - the spheres have no independent degrees of freedom of their own; only the clump as a whole has its own position, orientation, linear velocity and angular velocity. This is how BrixDEM represents non-spherical grains without giving up sphere-sphere contact detection, which is cheap and robust.

Mass properties

The clump's mass, center of mass and inertia tensor come from the geometric union of the N spheres - not from the naive sum, which would count the overlap region twice: the volume is integrated numerically (voxelization) over the composite shape and multiplied by the material density. The center of mass is the volume-weighted average, and the inertia tensor uses the parallel-axis theorem to shift each sphere's own inertia to the clump's center of mass:

equation 3.54(3.54)

where dᵢ is the distance from sphere i's center to the clump's center of mass (the overlap between spheres is already accounted for by the voxelization). Alternatively, BrixDEM can compute mass/COM/inertia from the exact STL solid that generated the clump, when available (10.8) - more accurate for shapes with substantial overlap, at the cost of requiring a closed mesh.

Contact and integration

Each pebble takes part in contact detection normally, as if it were an ordinary spherical particle - against other particles, against other clumps, and against walls/STL. The difference is in the integration: instead of each pebble receiving its own linear and angular acceleration, the forces and torques from every contact of every pebble in the clump are summed and applied to the single rigid body (using each pebble's position relative to the center of mass as the torque's lever arm). The result is integrated once per clump, not once per pebble - the clump advances in time as the large, rigid particle it represents, and the spheres simply follow that kinematics, rigid with respect to each other.

Configuration in the GUI - How to build a clump (presets, from STL, or manual/CSV), the sphere limits per template, and the usage rules are in Chapter 10.8.

3.15 Liquid bridges

A small amount of liquid outside a grain can form a meniscus between contacting surfaces. Pressure difference and surface tension create attraction; relative motion can also dissipate energy through viscosity. BrixDEM uses the Mikami correlation for attraction and rupture distance, with optional normal and tangential resistance from the Nase viscous approximation. A bridge forms at contact, remains during separation and ruptures beyond its critical distance. Proximity alone, without prior bridge formation, does not mean that a bridge is active.

Geometry, units and conventions

Consider spheres with radii Rᵢ and Rⱼ, center distance d and bridge volume Vb [m³]. The capillary radius rp differs from the lubrication radius Rv:

r_p=\frac{2R_iR_j}{R_i+R_j},\quad R_v=\frac{R_iR_j}{R_i+R_j}=\frac{r_p}{2} ; h=\max(d-R_i-R_j,0),\quad \widehat V=\frac{V_b}{r_p^3},\quad\widehat h=\frac{h}{r_p}(3.55)

h [m] is the nonnegative gap: the capillary correlation uses h = 0 during overlap. For sphere–wall contact, use rp = Rv = R, the sphere radius. Surface tension γl is in N/m; θ is converted from degrees in the GUI to radians in the equations. Do not confuse γl with the surface-energy parameter of the JKR model.

Capillary force and rupture

The attractive-force magnitude follows the expression below. Its direction points toward the other surface, with an equal and opposite reaction on the other body.

F_{\mathrm{cap}}=\pi r_p\gamma_l\,\max[\exp(A\widehat h+B)+C,0](3.56)

For sphere–sphere contact, the coefficients depend on dimensionless bridge volume:

A=-1.1\widehat V^{-0.53} ; B=(-0.34\ln\widehat V-0.96)\theta^2-0.019\ln\widehat V+0.48 ; C=0.0042\ln\widehat V+0.078(3.57)

For sphere–wall contact, with V̂ = Vb/R³ and ĥ = h/R, use:

A=-1.9\widehat V^{-0.51} ; B=(-0.016\ln\widehat V-0.76)\theta^2-0.12\ln\widehat V+1.2 ; C=0.013\ln\widehat V+0.18(3.58)

Do not add an external cos θ factor: angular dependence is already included in the published coefficients, including θ². An existing bridge stops acting when h exceeds its critical distance:

h_{c,pp}=(0.62\theta+0.99)\widehat V^{0.34}r_p ; h_{c,pw}=(0.22\theta+0.95)\widehat V^{0.32}R(3.59)

LITERATURE[32] — Mikami, Kamiya & Horio (1998), force and rupture correlations, Eqs. 22–25. Unequal spheres use rₚ; walls use their own correlation.

Calculated example. Two equal spheres of radius 1 mm, γl = 0.072 N/m, θ = 0° and V̂ = 0.002 have Vb = 2 × 10⁻¹² m³. At h = 0 the equation gives Fcap ≈ 4.23101 × 10⁻⁴ N; rupture occurs at hc ≈ 1.19670 × 10⁻⁴ m. These are mathematical reference values of the correlation, not experimental measurements.

Viscous contribution and minimum separation

When enabled, viscous resistance uses dynamic viscosity η [Pa·s]; the GUI accepts mPa·s, with 1 mPa·s = 10⁻³ Pa·s. S₀ [m] regularizes nearly touching surfaces:

S=\max(h,S_0),\qquad c_n=6\pi\eta\frac{R_v^2}{S} ; c_t=6\pi\eta R_v\max\left[\frac{8}{15}\ln\left(\frac{R_v}{S}\right)+0.9588,0\right](3.60)

cn and ct have units N·s/m. Using relative surface velocity and unit normal n, viscous force opposes motion:

\mathbf{v}_t=\mathbf{v}_{rel}-(\mathbf{v}_{rel}\cdot\mathbf{n})\mathbf{n},\quad v_n=\mathbf{v}_{rel}\cdot\mathbf{n} ; \mathbf{F}_{vis}=-c_nv_n\mathbf{n}-c_t\mathbf{v}_t,\quad\mathbf{F}_{vis}\cdot\mathbf{v}_{rel}=-c_nv_n^2-c_t|\mathbf{v}_t|^2\leq0(3.61)

Tangential velocity includes surface rotation. The tangential contribution also produces torque through the contact lever arm. The nonnegative logarithmic factor prevents a tangential force that supplies energy outside the asymptotic approximation. Decreasing S₀ increases normal resistance near contact; it does not change static capillary force or rupture distance. At zero relative velocity, viscous force is zero.

LITERATURE[33] — Nase et al. (2001), near-contact viscous resistance. Minimum separation and the nonnegative bracket are implementation regularizations.

Liquid content, surface inventory and volume modes

GUI liquid content w is a percentage of dry-solid mass md, rather than void volume or mixture mass. Liquid density ρl [kg/m³] converts initial liquid mass to film volume:

w=100\frac{m_l}{m_d},\qquad V_{film,0}=\frac{w}{100}\frac{m_d}{\rho_l}(3.62)

Thus 1% means 10 g/kg dry solid. A sphere with R = 1 mm and ρs = 2,500 kg/m³ has dry mass 1.04720 × 10⁻⁵ kg; with water density 1,000 kg/m³ and 1% content, its initial inventory is 1.04720 × 10⁻¹⁰ m³. Film is a scalar available reserve, not a resolved three-dimensional coating or liquid absorbed inside the grain.

Prescribed volume: when both material contents are zero, the pair specifies volume per bridge through the relative value. Migration: when content is positive, liquid comes from films; the same relative value limits what each bridge can receive, subject to inventory and competing contacts. A wall receives liquid from the particle and has no donating inventory.

V_{b,\max}=\widehat V_{\max}r_p^3,\qquad0\leq V_b\leq V_{b,\max} ; V_{tot}=\sum_i V_{film,i}+\sum_{b\in pp}V_b+\sum_{b\in pw}V_b(3.63)

The sum counts each bridge once, including wall bridges. In a closed system with inventory, formation and rupture redistribute liquid between films and bridges; Vtot should remain constant within numerical tolerance. For an open system, include introduced and removed liquid. Prescribed mode does not represent a conserved moisture inventory: bridge count can change without a surface reserve supplying it. A limit V̂max = 0.05 means neither 5% liquid content nor a universal saturation boundary.

Coarse-graining and validity

For local geometric similarity, the adopted parameter scaling preserves dimensionless arguments of the correlation and makes interaction forces grow as k²:

R,h,S_0\mapsto k(R,h,S_0),\quad\gamma_l,\eta\mapsto k(\gamma_l,\eta),\quad V_b\mapsto k^3V_b ; F_{cap},F_{vis}\mapsto k^2(F_{cap},F_{vis}),\qquad T_{vis}\mapsto k^3T_{vis}(3.64)

These relations require the same angle, volume scaled by k³ and identical relative surface velocities. Preserving surface speed in a rotating initial condition requires ω to scale as 1/k; the input does not apply that conversion automatically. Enter physical properties in the GUI: do not manually multiply γl or η before applying CGF. Liquid content, liquid density and V̂max remain unchanged.

LITERATURE[38], [39] — Larijani, Magnanimo & Luding (2025) support parameter scaling for k² interaction forces. Chan & Washino (2018) use a different strategy, evaluating at original scale before an external multiplier; the implementations are not identical.

Physical validity. This is an isolated pendular-bridge model. It does not resolve continuous liquid, pore pressure, absorption, connected menisci or saturation; an accepted liquid-content input can fall outside that regime. In the present model, liquid inventory does not add mechanical mass/inertia to the solid. Inventory conservation and agreement with the correlation verify computation, but do not automatically validate a real material or coarse-grained bulk dynamics. Configuration and compatibility: 10.9; results: 16.11; numerical checks: Appendix H.

3.16 Breakage and fragmentation

A breakage description separates three questions: which body reaches fracture, what damage remains after a non-fracturing event, and how mass is distributed among products. Ab-T10 describes selection, damage and t10; the Particle Replacement Method (PRM) replaces the fractured particle with discrete fragments. PRM is a replacement strategy, not an independent mineral-strength law.

Event energy and partition between bodies

An event groups loading of a body by its contacts. Energy input is assigned contact work divided by body mass; it is not automatically nominal drop-weight energy mgh, machine kinetic energy, or the instantaneous “Impact energy” scalar. The implementation accumulates normal and, when selected, tangential work increments:

\Delta W_{i,c}=\chi_{i,c}\,[F_{n,c}^{+}\Delta\delta_c+c_s\Delta W_{t,c}],\qquad e_{evt}=\frac{W_{evt}}{m_i} ; \chi_{i,c}=\frac{k_c}{k_i+k_c},\qquad k_j=\frac{E_j}{1-\nu_j^2}(3.65)

Fn+ is compressive normal force, Δδ is the overlap change and ΔWt is tangential work computed during loading. cs is the tangential fraction included in body breakage. Partition χ uses each material’s Hertzian stiffness [Pa]; it is 1/2 for equally stiff bodies. Normal work is signed through the contact cycle and the event retains the loading peak; summing absolute increments would define a different energy.

Fracture decision can use the event peak/end or resistance crossing during loading. Dialog choices and early decision under quasi-static compression are detailed in 10.10. Energy reached at the decision affects t10 and progeny; different protocols therefore do not automatically share calibrated parameters.

LITERATURE[36], [37] — Fracture energy, elastic partition and loading events. Event discretization and replacement timing must be kept consistent when reproducing a calibration.

Ab-T10: size, probability and accumulated damage

Let d [m] be representative diameter; dry clumps use volume-equivalent diameter. S [kg/J], dref [m] and emin,ref [J/kg] define the size-dependent selection coefficient and cutoff:

e_{min}(d)=e_{min,ref}\frac{d_{ref}}{d},\qquad a(d)=S\frac{d}{d_{ref}}(3.66)

For new particles of identical size under one impact with specific energy e [J/kg], breakage probability is P₁. Across successive events, net exposure sums the part above the cutoff:

P_1(e,d)=1-\exp[-a(d)\max(0,e-e_{min})] ; P_N=1-\exp\left[-a(d)\sum_{j=1}^{N}\max(0,e_j-e_{min})\right](3.67)

P is a population fraction, not the fraction of mass fragmented within one particle. Each body receives one initial sampled threshold. When an event does not fracture it, damage reduces remaining resistance:

T_0=e_{min}-\frac{\ln(1-u)}{a(d)},\quad u\sim U(0,1) ; T_{new}=\max[e_{min},\ T_{old}-\max(0,e_{evt}-e_{min})]\quad(e_{evt}<T_{old})(3.68)

The body fractures when eevt reaches its remaining threshold. An event below emin does not lower that threshold. A new independent resistance is not sampled at every collision. Optional inheritance applies a fraction of prior parent exposure to children, subject to the model floor; zero inheritance does not disable the body’s own damage accumulation.

LITERATURE[34], [35] — Vogel–Peukert energy/size selection and Shi–Kojovic size–energy relation; the equations show BrixDEM parameterization and accumulation.

Ab-T10: t10, reference energy and curve preview

GUI parameter A is a percentage; M = A/100 is its fractional form. At fracture, net energy can include prior accumulated exposure H or only the breaking event:

q_{10}=M[1-\exp(-a(d)e_{net})],\quad M=\frac{A_{\%}}{100} ; e_{net}=\max(0,e_{evt}-e_{min})+H,\quad H\in\{T_0-T_{old},0\}(3.69)

H = T₀ − Told for accumulated energy and H = 0 for event-only energy. A changes the t10 limit without changing P₁. For a new population, the single-impact preview gives:

e_{50}=e_{min}+\frac{\ln2}{a(d)},\qquad e_{90}=e_{min}+\frac{\ln10}{a(d)}(3.70)

e₅₀ is the energy at which 50% of a new population breaks, not deterministic strength of each body. At this point, single-impact predicted t10 is A/2. Prior damage and size changes invalidate this direct interpretation. The alternative JK b and xref representation converts as:

S=\frac{b_{JK}d_{ref}}{3600x_{ref}},\qquad1\ \mathrm{kWh/t}=3600\ \mathrm{J/kg}(3.71)

xref and dref must use the same length unit. The 3,600 factor converts kWh/t to J/kg; entering b as if it were S changes the energy scale. The preview checks parameterization but is not experimental validation.

Progeny and PRM replacement

t10 is mass passing an aperture equal to one tenth of parent size. Distinguish t10, breakage probability and child count. The incomplete-beta family relates t10 to other tn markers:

q_{10}=\frac{m(x<d_p/10)}{m_p},\quad t_{10}(\%)=100q_{10} ; q_n=I_{q_{10}}(\alpha_n,\beta_n)=\frac{\int_0^{q_{10}}z^{\alpha_n-1}(1-z)^{\beta_n-1}\,dz}{B(\alpha_n,\beta_n)}(3.72)

αn and βn are dimensionless, material-dependent parameters. The solver interpolates points logarithmically and enforces anchor Q(0.1) = q₁₀. The Gaudin–Schuhmann alternative uses the same anchor with a power curve:

Q(r)=r^p,\quad p=\frac{\ln q_{10}}{\ln0.1},\qquad r=\frac{x}{d_p}\in(0,1)(3.73)

A finite set of fragments and, where applicable, fines representatives discretize the continuous distribution. Maximum fragments controls resolution and cost. A rebreakage cutoff does not guarantee that no smaller child can be born; see the Ab-T10/PRM rule in 10.10. Mass accounting includes fines:

m_p=\sum_j m_j+m_{fines},\qquad Q(0.1)=q_{10}(3.74)

Replacement must conserve mass, with momentum transfer and initial separation verified. Collision-peak energy should not be interpreted as energy entirely recovered in child motion. Agreement of mass/t10 verifies discretization; experimental PSD comparison assesses material representation.

LITERATURE[36], [37] — Incomplete-beta distributions and stochastic particle replacement; conservation rules and evidence are in Appendix H.

From theory to configuration. Chapter 10.10 explains fields and presets. Calibrate energy, resistance and PSD under the same test protocol (18.4 and 18.6), check the GUI’s suggested dt (12.6), and examine quantitative results and verification limits in Appendix H. Compatibility with particle shape, liquid and CGF is collected in Chapter 28.