A Circuit-Inspired Framework for Manufacturing Systems

Abstract

This paper presents a framework for modeling and analyzing manufacturing systems by drawing analogies with electrical circuits. The approach, termed the Manufacturing Circuit Model, provides a physics-based perspective that complements traditional methods. By mapping manufacturing elements to circuit components, the framework enables the application of circuit theory to quantify system dynamics, size buffers, and translate operational change into cost. Key concepts include the Manufacturing Power Factor (MPF), damping-based buffer sizing, and frequency-domain analysis of production flow. Every parameter in the model traces to a measurement or a chosen policy value, so the framework carries no free fitting constants.

1. Introduction

Manufacturing systems are complex networks of interconnected processes, material flows, and information exchanges. Analyzing and optimizing these systems is critical for improving efficiency, reducing waste, and enhancing responsiveness. Traditional approaches such as queuing theory, simulation, and discrete-event modeling have provided valuable insights but often lack a cohesive physics-based framework that captures the dynamic interplay of system components.

This paper draws a direct analogy between manufacturing systems and electrical circuits. The resulting Manufacturing Circuit Model gives a compact, measurable way to reason about how a production line responds to changes in demand.

2. Literature Review

Existing manufacturing system analysis methods include:

• Queuing theory: models waiting lines and congestion but focuses on steady-state averages rather than transient response.
• Discrete-event simulation: powerful for complex interactions but lacks a general closed-form framework and can be computationally intensive.
• Lean manufacturing: emphasizes waste reduction and flow but lacks a quantitative approach to system-level dynamics.
• System dynamics: uses feedback loops and delays to model behavior; the circuit model is a specialization of this idea with named, measurable elements.
• Factory physics: provides quantitative relationships such as Little's Law but stops short of a component-level circuit analogy.

The Manufacturing Circuit Model builds on these foundations with a framework that is intuitive, mathematically explicit, and grounded in measurement.

3. The Manufacturing Circuit Model

Base units throughout: parts $\text{P}$ and seconds $\text{s}$.

Table 1: Circuit-manufacturing analogies

Demand enters as a current source, not a voltage source: the customer sets a rate, and the line either meets it or the buffer and backlog absorb the difference. Because demand is a current source, the frequency-domain quantities below come from the transfer functions of the 3.3 state equations, not from a series-RLC impedance triangle: in this model $R_{\text{eff}}$ sits in the stiffness term, not the dissipative one.

3.1 The three elements, grounded

• Resistance $R_{\text{eff}} = t_c / (m\,D)$ in s/P. The effective cycle time of a station: nominal cycle time $t_c$ divided by the number of parallel machines $m$ and the duty cycle $D$ (fraction of calendar time the station can run). Its reciprocal $\mu = 1 / R_{\text{eff}}$ in P/s is the station's maximum rate.
• Capacitance $C = Q_{\max}$ in P. The physical buffer between two stations. The buffer's potential is the dimensionless fill fraction $V_C = Q / Q_{\max}$, so $C = Q / V_C$ has units of parts. The capacitor relation is exact: $\dot Q = I_{\text{in}} - I_{\text{out}}$.
• Inductance $L = R_{\text{eff}}\,\tau_L$ in s²/P. A station cannot change its output rate instantly: WIP has to be repositioned, tooling changed, staffing and line balance adjusted. The lag $\tau_L$ in s is the measured time for output to reach 63% of a new level after a sustained demand step. The stored quantity $\tfrac{1}{2} L I^2 = \tfrac{1}{2} R_{\text{eff}}\,\tau_L\,I^2$ has units of parts and represents committed in-process work that will complete even if releases stop.

3.2 Circuit laws, with saturation

Throughput law (a saturating resistor). While the buffer has headroom, throughput is set by the fill fraction and the effective cycle time:

Units: dimensionless divided by s/P gives P/s.

What the throughput law represents. It is a release policy, not machine physics. A station left alone runs at its full rate $\mu = 1/R_{\text{eff}}$ whenever a part is waiting, which would make raw output $\mu$ for any $Q > 0$, not proportional to $Q$. Instead $I = V_C / R_{\text{eff}}$ is read as a WIP-proportional pull rule: the station is authorized to pull and release work at a rate proportional to how full its inbound buffer is, the linear form of the CONWIP and kanban principle that a stage should not outrun its downstream signal. This is what makes the linear-in-$Q$ form legitimate: it is a chosen control law, not a claim about the equipment. The release gain in this policy is set to $1/R_{\text{eff}}$ so the authorized rate meets true capacity exactly at a full buffer, and the rest of the paper uses that setting, so $\mu = 1/R_{\text{eff}}$ throughout and the saturation clause below still binds at true capacity. Running a deliberately throttled gain $g < 1/R_{\text{eff}}$ is possible but is a separate design choice that lowers the effective capacity to $g$; it is not considered here. So $\omega_n$, $\zeta$, $\phi$, and $A_Q$ below describe a controlled line, not an open-loop one.

When demand exceeds capacity, $\lambda \ge \mu = 1 / R_{\text{eff}}$, the fill fraction pins at 1, throughput clamps at $\mu$, and unmet demand accumulates as backlog $B$ with $\dot B = \lambda - \mu$. A smooth approximation across the transition is the parallel-conductance form

with all three terms in s/P. This replaces the linear $V = I R$, which does not close dimensionally and does not capture that throughput saturates at capacity.

Conservation of flow (Kirchhoff's current law). At every node, units are conserved:

3.3 Second-order dynamics

Couple the buffer (capacitor) to the throughput lag (inductor). With $Q$ the buffer level and $I$ the station output:

Eliminating $I$ gives a damped second-order system in the buffer level:

with natural frequency and damping ratio

Both expressions unit-check. Every parameter is fixed before the model runs: $R_{\text{eff}}$ from a cycle-time study (and used as the release gain, 3.2), $C$ from the buffer, $\tau_L$ from an output step test. There are no free constants to fit.

4. Key Concepts and Applications

4.1 Manufacturing Power Factor (MPF)

Transforming the 3.3 state equations, with demand $\lambda$ as the current source and station output $I$ as the response, gives the demand-to-throughput transfer function

a second-order low-pass with the $\omega_n$ and $\zeta$ of 3.3. At a demand frequency $\omega$, throughput lags demand by the phase angle

Numerator and denominator both carry units of $\text{s}^{-2}$, so $\tan\phi$ is dimensionless. Define

Interpretation. Split the throughput response at frequency $\omega$ into a part in phase with demand and a part in quadrature. The in-phase part, scaling with $\cos\phi$, is flow that serves demand. The quadrature part, scaling with $\sin\phi$, fills the buffer over one half of the cycle and draws it back down over the other, contributing zero net output while still consuming labor, machine time, and carrying cost. MPF is the fraction of the throughput response that tracks demand, on a 0 to 1 scale: it is 1 when output is in phase and falls to 0 at $\omega = \omega_n$, where the response is pure buffer exchange. For $\omega \ge \omega_n$ the raw $\cos\phi$ is negative, meaning output moves against demand, and MPF is clipped to 0; the meaningful quantity in that regime is the phase lag $\phi$ itself, not its cosine. MPF reaching 0 at $\omega_n$ is the frequency-domain form of the warning in 4.2: do not let the dominant demand variation sit at the natural frequency.

The lever is $\tau_L$. For $\omega < \omega_n$, cutting the throughput adjustment lag $\tau_L$ lowers $\tan\phi$ and raises MPF, and at the same time raises $\omega_n$ and $\zeta$ (4.2). Changeover reduction (SMED), flexible staffing, and smaller transfer batches all cut $\tau_L$. The lag has a floor: as $\tau_L \to 0$ the model collapses to the first-order buffer lag $\phi \to \arctan(\omega\,Q_{\max}R_{\text{eff}})$, set by the buffer time constant $Q_{\max}R_{\text{eff}}$. A larger buffer raises damping (4.2) but also raises this residual lag, so buffer size trades demand tracking against stability instead of improving both.

WIP swing. The buffer response to demand follows from $\dot Q = \lambda - I$:

with $A_Q$ in seconds. A demand component of amplitude $\Delta\lambda$ at frequency $\omega$ drives a buffer oscillation of amplitude $\Delta Q = \Delta\lambda\,A_Q(\omega)$. At low frequency $A_Q \to Q_{\max}R_{\text{eff}}$, so a slow demand change of size $\Delta\lambda$ shifts the fill fraction by $\Delta\lambda/\mu$, matching $V_C^{\ast} = \lambda/\mu$ in 4.3. Near $\omega_n$, $A_Q$ rises above that floor whenever $\zeta < \tfrac{1}{2}\sqrt{1 + \sqrt{2}} \approx 0.777$: this is the bullwhip. The threshold is not $1/\sqrt{2}$, because the zero at $s = -1/\tau_L$ in $Q/\lambda$ raises it; see 4.2. Meeting the 4.2 buffer rule keeps $A_Q$ at or below its low-frequency floor across the demand band, and well above $\omega_n$ the buffer filters the swing, so $A_Q$ falls below the floor.

4.2 Buffer sizing by damping

An earlier version of this framework prescribed operating at resonance, $\omega_n = \omega_{\text{takt}}$. That is wrong. Near $\omega_n$ a lightly damped line amplifies buffer swings (the bullwhip or hunting mode), which is the behavior to suppress, not to seek.

The design target is a demand-to-WIP response $A_Q(\omega)$ (4.1) with no peak above its low-frequency value. The buffer transfer function $Q/\lambda$ carries a zero at $s = -1/\tau_L$ that lifts its high-frequency side, so the flat-response condition is not $\zeta \ge 1/\sqrt{2}$. Writing $|Q/\lambda|^2$ as $(x + 4\zeta^2)/\bigl((1-x)^2 + 4\zeta^2 x\bigr)$ with $x = (\omega/\omega_n)^2$, its slope at $x = 0$ is $1 + 8\zeta^2 - 16\zeta^4$, which stays positive (the response rises above its DC value) until

Units: s divided by s/P gives P. The plain $\zeta \ge 1/\sqrt{2}$ threshold still applies, but to the demand-to-throughput response $I/\lambda$, which is a pure second-order low-pass with no zero: size to it when the goal is a smooth output rate, and to the $(1 + \sqrt{2})$ rule when sizing the buffer against WIP swing. For a monotonic WIP settle with no overshoot at all, require $\zeta \ge 1$:

$\omega_n$ is still useful: it is the frequency of demand variation the line is least able to reject. Keeping $\zeta$ above $\tfrac{1}{2}\sqrt{1 + \sqrt{2}}$ ensures buffer fluctuations in that band are not amplified.

4.3 Buffer fill and drain timing

These quantities are kinematic (pure flow balance) and hold regardless of the dynamic model. Let $I_{\text{up}}$ and $I_{\text{down}}$ be the upstream and downstream throughputs.

Time to fill if the downstream station stops:

Time to drain if the upstream station stops:

Net rate of change of the buffer level:

In subcritical steady state both stations run at the demand rate, $I_{\text{up}} = I_{\text{down}} = \lambda$, so $dQ/dt = 0$ and the buffer settles at a fill fraction equal to utilization:

Minimum buffer to cover a planned upstream outage of duration $t_{\text{out}}$ without starving downstream:

Take the larger of $Q_{\min}$ (outage coverage) and the damping requirement $(1 + \sqrt{2})\,\tau_L / R_{\text{eff}}$ from 4.2.

4.4 Duty cycle and effective resistance

Availability lives inside $R_{\text{eff}}$. For a station with nominal cycle time $t_c$ in s/P, $m$ parallel machines, and duty cycle

the effective resistance and maximum rate are

Typical values of $D$: one 8-hour shift, $8/24 = 0.33$; two shifts, $16/24 = 0.67$; three shifts, $1.0$; weekend-only, $16/168 = 0.095$.

The duty cycle also scales the dynamic response. Since $R_{\text{eff}} \propto 1/D$,

Lower availability slows the line's response to demand shifts as well as its rate. This replaces the earlier linear claim $\omega_{\text{eff}} = \omega_n\,D$; the square-root scaling is what the state equations give.

4.5 Batch processing

Batch processes (ovens, soak tanks, leak tests) hold units for a fixed time $\tau$ in s, then release them. They behave as a pure delay plus a moving average.

Normalized transfer function, from input rate to output rate, with unit DC gain:

The magnitude $\lvert\operatorname{sinc}(\omega\tau/2)\rvert$ shows the batch stage passing low-frequency demand trends and attenuating high-frequency variation, a low-pass filter with its first null at $\omega = 2\pi/\tau$.

Capacity, with batch size $b$ in P and $N$ parallel processors, kept separate from the frequency response:

Staggering for continuous flow. To make a batch stage approximate continuous flow at takt period $T_{\text{takt}}$, run

overlapping batches, each offset by $\tau / n$. The buffer feeding the stage must hold at least $n\,b$ units so a full batch is always ready.

Worked example. Oven with $\tau = 84{,}807$ s, $b = 4$ P, $N = 2$. At a takt period of $T_{\text{takt}} \approx 17{,}300$ s, $n = \operatorname{round}(4.9) = 5$ overlapping batches, feed buffer at least $20$ units, and $\mu_{\text{batch}} = 2 \cdot 4 / 84{,}807 \approx 9.4 \times 10^{-5}$ P/s, about 8 units per day.

Integration.

• Duty cycle: limited availability of a batch station lengthens the effective stagger interval and forces schedule adjustments to protect downstream flow.
• Frequency domain: a batch stage is a low-pass filter, so it smooths high-frequency demand variation before it reaches downstream stations.
• MPF: aligning batch starts with takt and holding $n$ near $\tau / T_{\text{takt}}$ minimizes the WIP that circulates through the stage, reducing its contribution to the buffer swing $A_Q$ of 4.1.

4.6 Financial translation

The recoverable cost of dynamic inefficiency is the cost of WIP that circulates without net output, plus the cost of driving throughput up and down to produce that circulation. Both come from the quantities in 4.1, and both are sized from measured parameters. The earlier version of this section split cost by a power-factor identity, $K_{\text{VA}} = K\cos\phi$ with a reactive remainder; that identity came from a series-RLC impedance triangle the model does not produce, and its pieces did not sum to the whole. It is replaced below.

Take a demand band with dominant frequency $\omega$ and amplitude $\Delta\lambda$ about the mean rate. The buffer swings with amplitude $\Delta Q = \Delta\lambda\,A_Q(\omega)$. The unavoidable part is the static fill shift $\Delta\lambda\,Q_{\max}R_{\text{eff}}$; the remainder,

is oscillating WIP that a faster-adjusting, better-damped line would not hold. Below the corner the buffer amplifies the swing and this term is positive; well above $\omega_n$, $A_Q$ drops below the floor because the buffer filters the swing, the term clamps to zero, and the model predicts no excess carrying cost from high-frequency demand components. That is a checkable prediction.

Carrying cost. A swing about a fixed mean does not by itself raise mean WIP: the half-cycle above the mean cancels the half-cycle below. The cost enters through safety stock. To keep the trough of the swing from starving the downstream station, the mean buffer level has to be lifted by about $\Delta Q_{\text{exc}}$, and that raised mean is carried the whole time. With $h$ the holding cost per part per unit time (capital, floor space, handling, obsolescence risk),

If the binding constraint is physical buffer size rather than carrying cost, the same $\Delta Q_{\text{exc}}$ drives a one-time capital cost for the larger buffer instead.

Ramp cost. Producing the swing forces throughput itself to rise and fall by $\Delta I = \Delta\lambda\,\lvert I/\lambda(j\omega)\rvert$. With $c_{\text{ramp}}$ the measured incremental cost per unit of throughput swing (extra changeovers, flex staffing, expediting) and $N_{\text{cyc}}$ demand cycles in the period,

Return on a lag-reduction project. Lowering $\tau_L$ raises $\omega_n$ and $\zeta$ and pulls $A_Q(\omega)$ toward its floor, shrinking both $\Delta Q_{\text{exc}}$ and $\Delta I$. With subscripts 0 and 1 for before and after,

Illustrative example. Numbers are round, not measured. A line with $R_{\text{eff}} = 100\ \text{s/P}$, $Q_{\max} = 1000\ \text{P}$, and $\tau_L = 5\times10^{5}\ \text{s}$ has $\omega_n \approx 4.5\times10^{-6}\ \text{s}^{-1}$ and $\zeta \approx 0.22$. A monthly demand swing, $\omega \approx 2.4\times10^{-6}\ \text{s}^{-1}$, amplitude $\Delta\lambda = 0.002\ \text{P/s}$, gives $A_Q \approx 2.1\times10^{5}\ \text{s}$ against a floor $Q_{\max}R_{\text{eff}} = 1.0\times10^{5}\ \text{s}$, so $\Delta Q \approx 420\ \text{P}$ with about $220\ \text{P}$ of it avoidable. A SMED and transfer-batch program that halves $\tau_L$ to $2.5\times10^{5}\ \text{s}$ raises $\omega_n$ to $\approx 6.3\times10^{-6}\ \text{s}^{-1}$ and $\zeta$ to $\approx 0.32$, dropping $A_Q$ to $\approx 1.3\times10^{5}\ \text{s}$ and the avoidable WIP to about $64\ \text{P}$. The annual carrying saving is $(220 - 64)\,\text{P}\times h \times 8760\ \text{h/yr}$ from the reduced safety stock; the ramp saving from the smaller $\Delta I$ adds to it. Dividing the sum by project cost gives first-year ROI, computed from a $\tau_L$ change rather than an assumed MPF gain.

The example also shows the diagnostic side of 4.2. Both states have $\zeta$ far below the $0.777$ threshold ($0.22$, then $0.32$), and the buffer rule $Q_{\max} \ge (1 + \sqrt{2})\,\tau_L / R_{\text{eff}}$ calls for $\approx 12{,}000\ \text{P}$ before the project and $\approx 6{,}000\ \text{P}$ after, against an actual buffer of $1{,}000\ \text{P}$. The model is flagging the line as badly under-damped: halving $\tau_L$ is a real improvement but nowhere near enough, and either the buffer or the lag has to move much further before the bullwhip is gone.

Fit with standard tools. $K_{\text{WIP}}$ and $K_{\text{ramp}}$ are ordinary cost lines, so their projected reductions drop into net present value, payback, and activity-based costing unchanged. What the model adds is a way to size them from $\tau_L$, $R_{\text{eff}}$, $Q_{\max}$, and the demand spectrum.