Explicit vs implicit solver: forward Euler load steps and Newton-Raphson iterations, step by step
Loaded in three steps of ΔF = 1, the bar F(u) = u³ + 9u² + 4u ends at u = 0.456250 with the explicit (forward Euler) method, 16.7% past the exact 0.391026. Newton-Raphson with |R| < 10⁻² ends at 0.391107, after 2, 1 and 1 corrections. Change the force law, the load, the steps or the tolerance and every iteration is recomputed. Free, runs in your browser.
The defaults are the "Week 2 - Explicit and Implicit Analysis" challenge in Skill-Lync's LS-DYNA course. Skill-Lync read 262 public reports on it: 162 give the right explicit u3 and 209 the right implicit u3 in their text, but 44 reach step 1 with a modified-Newton step they call Newton-Raphson. The second tab shows what LS-DYNA's explicit solver actually does.
Yaw (2009) example the challenge copies · LS-DYNA Theory Manual R14 ch. 27, 37 · Keyword Manual R14 *CONTROL_IMPLICIT_SOLUTION · Bathe, MIT OCW · checked against a separate solution (26,538 checks at build) · Updated 22 Nov 2026
The challenge: three increments of ΔF = 1, |R| < 10⁻². A correction is one Newton update after the ΔF/K predictor; the predictor itself is iteration 0.
Uses the force law above. Central difference is explicit (LS-DYNA's solver); Newmark average acceleration is implicit, with Newton iterations inside every time step. The blue reference is RK4 with tiny steps.
Checks against u³ + 9u² + 4u, three increments of ΔF = 1 and |R| < 10⁻². The diagnoses are the slips counted in the public reports.
- Forward Euler ends at u = 0.456250 with F_int = 3.79345 against F_ext = 3: 16.7% off in u, and nothing in the method notices.
- Full NR ends at u = 0.391107 (exact 0.391026), 2, 1, 1 corrections per step, 4 in all.
- Forward Euler needs 37 increments to stay within 1% of the exact u at F = 3; carrying the residual into the next increment needs 4.
Every step, every method
| Step | F_ext | Exact u | Forward Euler | Euler + residual | Full NR | Modified NR | Initial stiffness |
|---|---|---|---|---|---|---|---|
| 1 | 1 | 0.177617 | 0.250000 +41% | 0.250000 +41% | 0.177661 +0.025% · 2 corr. | 0.176434 −0.67% · 21 corr. | 0.176434 −0.67% · 21 corr. |
| 2 | 2 | 0.296159 | 0.365108 +23% | 0.298561 +0.81% | 0.296504 +0.12% · 1 corr. | 0.295542 −0.21% · 3 corr. | ✗ 0.489522 +65% · 50 corr. |
| 3 | 3 | 0.391026 | 0.456250 +17% | 0.399883 +2.3% | 0.391107 +0.021% · 1 corr. | 0.391394 +0.094% · 2 corr. | stopped |
Every iteration: full Newton-Raphson
| Step | Iter. | K used | Δu or δu | u | F_int(u) | R = F_int − F_ext | |R| (force) | < tol? |
|---|---|---|---|---|---|---|---|---|
| 1 | 0 (predictor) | 4.00000 | 0.25000 | 0.250000 | 1.57813 | 0.57813 | 0.57813 | no |
| 1 | 1 | 8.68750 | −0.06655 | 0.183453 | 1.04288 | 0.04288 | 0.04288 | no |
| 1 | 2 | 7.40312 | −0.00579 | 0.177661 | 1.00032 | 0.00032 | 0.00032 | yes |
| 2 | 0 (predictor) | 7.29258 | 0.13713 | 0.314786 | 2.18215 | 0.18215 | 0.18215 | no |
| 2 | 1 | 9.96343 | −0.01828 | 0.296504 | 2.00332 | 0.00332 | 0.00332 | yes |
| 3 | 0 (predictor) | 9.60082 | 0.10416 | 0.400662 | 3.11174 | 0.11174 | 0.11174 | no |
| 3 | 1 | 11.69351 | −0.00956 | 0.391107 | 3.00093 | 0.00093 | 0.00093 | yes |
What does "explicit" mean in the Week 2 challenge?
Forward Euler load stepping: add the load in increments and move the bar by ΔF/K each time, with the stiffness K taken at the start of the increment. There is no check that the bar is in equilibrium afterwards. For the challenge's bar:
The bar stiffens as it stretches, so using the starting stiffness for the whole increment always overshoots. After three steps the internal force is well above the load:
u3 is 16.7% larger than the exact 0.391026, and nothing in the method tells you. The error falls about as 1/n: this bar needs 37 increments to get within 1% and 354 for 0.1%. Feeding the unbalanced force of each step into the next increment (the "Euler + residual" column) needs 4 and 13. The chart under the calculator plots all three against the number of increments.
How many Newton-Raphson iterations does each step need?
Two in step 1, one in step 2 and one in step 3, with |R| < 10⁻². The implicit method starts each step with the same ΔF/K predictor, then corrects it until the residual is small enough:
Step 1, with K = 3u² + 18u + 4 evaluated at each new u:
Step 2 starts from the converged 0.177661: predictor 0.314786 (R = 0.1822), one correction to 0.296504 (R = 0.00332). Step 3: 0.400662 (R = 0.1117), then 0.391107 (R = 0.00093). The full table above lists every row.
The answer depends on what you call "converged". With the same 10⁻² on a relative displacement |δu|/|u| (the idea behind LS-DYNA's DCTOL) the steps take 3, 2, 2 corrections; on an energy ratio (ECTOL's idea) 2, 2, 1; on |R|/F_ext 2, 1, 1. "Only two Newton-Raphson iterations in each step" is a sentence from the source example, about a different bar.
Newton's convergence is quadratic close to the root: each residual is roughly a constant times the square of the last one.
In step 1 the ratio is 0.128 then 0.174 against the predicted 0.179: the residual goes from 0.5781 to 0.0429 to 3.2e-4. The residual chart in the calculator shows it per step.
Is modified Newton or initial stiffness cheaper?
Per iteration, yes: they reuse one stiffness instead of forming a new one each time. But they need many more iterations and lose the quadratic convergence (Bathe, MIT RES.2-002 lecture 10):
On this bar, modified Newton needs 21, 3 and 2 corrections. Initial stiffness, with K(0) = 4 throughout, does not converge at all in step 2: the error is multiplied by about
every iteration, and a factor bigger than 1 in size grows. A stiffening bar is the worst case for a stiffness frozen at the start. LS-DYNA's implicit solver defaults to BFGS updates (NSOLVR = 12), which sit between the two, and becomes full Newton with ILIMIT = 1 (keyword manual R14, *CONTROL_IMPLICIT_SOLUTION, remark 3).
What happens when the load passes a limit point?
Load control fails. Pick the softening preset:
Asked for 3.5 in 7 increments, every method is fine up to 3. At step 7 no equilibrium exists on the loading path. Full Newton "converges" to u = −2.3443, a root on another branch, and the calculator flags it; modified Newton and initial stiffness run away. Forward Euler prints u = 1.1627 as if nothing happened. Following a structure past its peak needs arc-length control; in LS-DYNA that is NSOLVR = 12 with ARCMTH = 3.
Is this what LS-DYNA explicit and implicit actually do?
No. The challenge is static: no mass, no time, and its "explicit" is a load-stepping scheme. LS-DYNA's explicit solver integrates the equations of motion in time with the central difference method, using a diagonal mass matrix, and is stable only below a critical time step (Theory Manual R14, ch. 27):
The second tab gives the same spring a mass m = 1 and a sudden load of 3. At rest K = 4, so Δt_cr = 1. But the bar swings to u = 0.707, twice its static 0.391, where K is 18.22 and Δt_cr drops to 0.4686. Δt = 0.4 stays bounded. Δt = 0.44 blows up at t = 7.92, though it is below 0.4686: the coarse run overshoots into stiffer states, and its own Δt_cr falls below 0.44 at t = 2.64. The safe step is set by the stiffest state the run reaches, not by the starting one.
The theory manual's damped limit is smaller:
This page's central difference takes the damping velocity centred (Bathe, MIT 2.092 lecture 16, eq. 8), and for that form the limit stays 2/ω (derived here from the root λ = −1 of its characteristic equation), so treat 2/ω as an upper bound. For real meshes, the element step comes from the wave speed and element size: the explicit time step calculator does that.
LS-DYNA's implicit dynamics uses Newmark; with γ = 1/2 and β = 1/4 "the method becomes the trapezoidal rule and is energy conserving" (Theory Manual, 37.4.1). Bathe's lecture 18 calls it "unconditionally stable" (for linear problems): Δt is chosen for accuracy, not stability. Each time step is a small nonlinear static problem solved by Newton:
Stable is not accurate. For a linear spring both schemes keep the amplitude but change the period (derived here from their characteristic equations, checked against the eigenvalues of each amplification matrix):
With ω = 2 (period 3.1416), Δt = 0.1 gives 3.1363 with central difference and 3.1520 with Newmark; Δt = 0.5 gives 3.0000 and 3.3879.
LS-DYNA's defaults on *CONTROL_IMPLICIT_SOLUTION (keyword manual R14) are DCTOL = 0.001 on displacement, ECTOL = 0.01 on energy and RCTOL = 10¹⁰ on the residual force, which leaves the force check effectively off. The challenge's |R| < 10⁻² is a force check; Bathe's lecture 10 gives RTOL = 0.01 as typical for a force criterion, relative to a reference force.
Where does the Week 2 problem come from?
From Louie L. Yaw's "Nonlinear Static - Explicit and Implicit Analysis Example" (Walla Walla University, draft 27 May 2009). Yaw's bar is f = u + u³/3, k = 1 + u², with the same three steps of ΔF = 1 and |R| < 10⁻²; the challenge swaps in u³ + 9u² + 4u. The sentences many reports carry ("three equal load increments (or steps) of ΔF = 1 unit each are used to load the tension bar") are his. Only 2 of 252 reports name him. The Yaw preset reproduces his tables:
- Table 1 (explicit): u = 1.00, 1.50, 1.81 (1.0000, 1.5000, 1.8077 here). His f_int(u3) = 3.79 is computed from u3 rounded to 1.81; unrounded it is 3.777.
- Step 1 (implicit): 1.0 → 0.83333 → 0.81785 with R = 0.000197; here 1.00000 → 0.83333 → 0.81785.
- Step 2: his two corrections divide by 3.42508 and 3.02729, but k = 1 + u² at 1.41705 and 1.31033 is 3.00803 and 2.71696. Full Newton gives 1.28794 (R = 7.4e-5); his 1.29042 still meets |R| < 10⁻², so the table stays within tolerance. His step-2 text also writes "Fext = ΔF = 1" and then subtracts 2.
- Step 3: 1.61 in his table; 1.6110 here.
What did 262 student reports answer?
Skill-Lync read all 276 public reports whose title pairs "explicit" and "implicit" on 22 Nov 2026. 13 are a different problem (2D heat conduction) and 1 is an Euler integration project; 262 are this bar. 252 show working for u³ + 9u² + 4u in their text, 2 use other coefficients and 8 have no numbers in their text. The tables sort the 252 by the u3 their text gives, the best value first. 68 have no images, so for them the text is everything; numbers inside images aren't counted.
| Explicit u3 in the text | Reports | Examples |
|---|---|---|
| 0.4562 or 0.4563 (0.456250) | 118 | 1, 2, 3 |
| 0.456: right to 3 figures | 44 | — |
| 0.46: rounded to 2 decimals | 39 | 1, 2, 3 |
| 0.458: k(u2) written as 10.7178 | 4 | 1, 2, 3 |
| another value | 22 | 1, 2, 3 |
| no explicit u3 in the text | 25 | — |
| Total | 252 |
| Implicit u3 in the text | Reports | Examples |
|---|---|---|
| 0.3911 or 0.3912 (0.391107) | 180 | 1, 2, 3 |
| 0.3910: the exact root, a tighter tolerance | 9 | 1, 2, 3 |
| 0.391: right to 3 figures | 20 | 1, 2, 3 |
| another value | 9 | 1, 2, 3 |
| no implicit u3 in the text | 34 | — |
| Total | 252 |
- Modified Newton by accident. 44 reports write step 1's second correction with k(0.25) = 8.6875 instead of k(0.1835) = 7.4031 and get 0.1786 (1, 2, 3). It passes 10⁻², but it is not the Newton-Raphson step the text names. 3 stop step 1 at 0.1835, where |R| = 0.043 (1, 2, 3).
- Rounding carried. 76 write u2 = 0.2966 from 0.3148 − 0.0182; unrounded it is 0.296504 (1, 2, 3).
- The source's typo. 13 copy Yaw's step-2 line "Fext = ΔF = 1" while subtracting 2 (1, 2, 3). No report's text balances step 2 against 1, which would give u2 ≈ 0.196.
- Iteration counts. 15 repeat "only two Newton-Raphson iterations are required" in each step (1, 2, 3); for this bar it is 2, 1, 1. 5 start step 2's Newton from the explicit 0.3651 (1, 2, 3).
- Theory. 34 describe LS-DYNA's explicit solver as central difference and 52 mention a critical time step, but the hand calculation the challenge asks for has no mass and no time step: the dynamic tab is the missing half.
Counts as of 22 Nov 2026, n = 252 reports with working in text (262 on the challenge), from the text of each public report (Skill-Lync's own data, a starting point, not a grade).
- Explicit u3 = 0.4584: k(u2) slipped. k(0.3651) = 3(0.3651)² + 18(0.3651) + 4 = 10.9719, not 10.7178, so Δu3 = 0.0911 and u3 = 0.4563. 4 public reports wrote this (1, 2, 3).
- Explicit F_int(u3) = 3.7935 against F_ext = 3: forward Euler drifts 16.7% too far in u. That is the method, not a slip.
- Implicit u1 = 0.1786: the second correction used k(u⁽⁰⁾) = k(0.25) = 8.6875 instead of k(u⁽¹⁾) = k(0.1835) = 7.4031. That is a modified Newton step: R = 0.0066 passes 10⁻², but full Newton gives 0.177661. 44 public reports wrote this (1, 2, 3).
- Implicit u2 = 0.2966: right to rounding. 0.3148 − 0.0182 = 0.2966 carries the rounding; unrounded, Newton gives 0.296504 (0.2965). 76 public reports wrote this (1, 2, 3).
- Implicit u3 = 0.3911: right (0.391107, |R| = 0.00093 < 10⁻²).
- Step 1: 2 corrections, right.
- Step 2: 2 corrections; with |R| < 10⁻² it takes 1 (after the predictor, residuals 0.18215 → 0.00332).
- Step 3: 2 corrections; with |R| < 10⁻² it takes 1 (after the predictor, residuals 0.11174 → 0.00093).
- "Only two Newton-Raphson iterations in each step" is Yaw's sentence about his u + u³/3 bar. For u³ + 9u² + 4u at |R| < 10⁻² the steps take 2, 1, 1. 15 public reports wrote this (1, 2, 3).
What this calculator does not cover
- More than one degree of freedom. Everything here is one spring. With a stiffness matrix the ideas are the same, but the cost of forming and factorising K is what makes modified Newton and BFGS worth it.
- Line search and arc length. LS-DYNA adds a line search to every scheme and offers arc-length control; neither is implemented, the page only shows where plain load control stops.
- Hourglass, contact and element time steps. The dynamic tab is a single spring and mass. For a mesh, use the explicit time step calculator.
Knowing why an implicit run won't converge, and when an explicit step will blow up, saves days on real crash and forming models. The explicit and implicit challenge is part of Skill-Lync's LS-DYNA course.
Frequently asked questions
What is the difference between explicit and implicit analysis?
In this challenge, explicit means adding ΔF/K(u) each load step with no check, so the answer drifts. Implicit means correcting each step with Newton-Raphson until the residual F_int − F_ext is below a tolerance. In LS-DYNA, explicit is central-difference time integration and implicit solves equilibrium with Newton-type iterations.
What is the explicit answer for u³ + 9u² + 4u in three steps?
u = 0.25, 0.365108 and 0.456250. Each step adds ΔF/K with K = 3u² + 18u + 4 taken at the start of the step: 4, 8.6875 and 10.9719. The internal force at the end is 3.7935 against an applied 3, so u is 16.7% too large.
What is the implicit (Newton-Raphson) answer?
u = 0.177661, 0.296504 and 0.391107 with |R| < 10⁻². The exact roots are 0.177617, 0.296159 and 0.391026. Step 1 needs two corrections after the predictor (0.25 → 0.183453 → 0.177661); steps 2 and 3 need one each, four corrections in all.
Why does the explicit method drift from the exact solution?
It uses the stiffness at the start of each increment for the whole increment. The bar stiffens as it stretches, so each step overshoots and the error is never corrected. The error falls about as 1/n: with forward Euler this bar needs 37 increments to get within 1% and 354 for 0.1%.
Is the explicit method in this challenge what LS-DYNA explicit does?
No. The challenge's explicit is static load stepping with no mass and no time. LS-DYNA explicit integrates m ü = P − F in time with central differences and is stable only while Δt ≤ 2/ω_max. The second tab runs that on the same spring, with a mass, and shows it blow up.
What convergence tolerances does LS-DYNA implicit use by default?
In the R14 keyword manual, *CONTROL_IMPLICIT_SOLUTION defaults are DCTOL = 0.001 (relative displacement), ECTOL = 0.01 (relative energy) and RCTOL = 10¹⁰ (residual force, so effectively off). The default solver is NSOLVR = 12, BFGS updates; ILIMIT = 1 makes it full Newton.
Related on Skill-Lync
- Explicit time step calculator: the Courant step of shells and solids, mass scaling
- Hyperelastic curve fit calculator: Mooney-Rivlin, Yeoh and Ogden constants for the same course
- Consistent units calculator: kg-mm-ms, tonne-mm-s and others
- 2D heat conduction solver: explicit and implicit time schemes for the heat equation
- LS-DYNA courses compared
- Student project: Week 2 Explicit and Implicit Analysis
- All free engineering tools
Sources
- Louie L. Yaw, Nonlinear Static - Explicit and Implicit Analysis Example, Walla Walla University, draft 27 May 2009: the problem, Tables 1 and 2, the step-1 and step-2 iterations (read in full).
- Ansys / LSTC, LS-DYNA Theory Manual R14 (02/24/23): ch. 27 (central difference, eqs. 27.6 to 27.11; stability, Δt ≤ 2/ωmax, eq. 27.27; damped, eq. 27.28) and ch. 37 (implicit: full Newton and quasi-Newton schemes, BFGS default; convergence with DCTOL, ECTOL, RCTOL, eqs. 37.99 to 37.101; Newmark, 37.4.1).
- Ansys / LSTC, LS-DYNA Keyword User's Manual Vol. I R14 (07/31/23), *CONTROL_IMPLICIT_SOLUTION: NSOLVR 12, DCTOL 0.001, ECTOL 0.01, RCTOL 10¹⁰, ILIMIT 11, remarks 2 and 3 (arc length with ARCMTH = 3; ILIMIT = 1 is full Newton).
- K. J. Bathe, MIT OCW RES.2-002 (2010), lecture 10: full Newton-Raphson, modified Newton and the initial stress method, force tolerance RTOL typically 0.01. MIT 2.092 (2009), lectures 16 and 18: central difference, Δtcr = Tn/π = 2/ωn, the trapezoidal rule "unconditionally stable".
- Skill-Lync, 276 public reports with "explicit" and "implicit" in the title, read 22 Nov 2026. scripts/verify/explicit_implicit.py solves every case another way (60-digit Decimal iterations, bisection and Cardano roots, a brute-force increment scan, amplification-matrix eigenvalues, scipy DOP853 for the dynamics), 304 cases; the page checks itself against it at build time (26,538 checks).