Design specifications, sensitivity analysis and case studies

With the flowsheet converging reliably, the questions from the project team began to arrive. Carla, the project manager, wanted to know what reflux ratio the ester column needed to meet the 99.7 % purity with the 24 stages the contractor had chosen. Wren, the operations lead who would run the line, asked what would happen to purity if the reflux drifted. Nadia, the quality manager, asked how robust the purity was to changes in the decanter feed. Each question is a different kind of simulation study: a design specification, a sensitivity analysis and a set of case studies.

Design specifications

A simulator's unit models take inputs and calculate outputs. A column model takes a reflux ratio and calculates the product purity. But the engineer often knows the purity required and wants the reflux. A design specification turns the problem around. It names a target, such as the mole fraction of light material in the bottoms equal to 0.003, and a manipulated variable, such as the reflux ratio, with bounds. The simulator then adjusts the manipulated variable until the target is met. The degrees of freedom are unchanged: one specification has been traded for another.

In a sequential-modular simulator, a design specification is a loop around part of the flowsheet. The simulator guesses the manipulated variable, calculates the units, compares the result with the target, and updates the guess, usually by a secant or Newton method. When a design specification sits inside a recycle, there are nested loops, or the simulator converges the specification and the tears together, which a quasi-Newton method does well.

The choice of the target function matters. Bottoms impurity falls roughly exponentially as reflux rises, so \(x_{B}\) itself is a strongly curved function of \(R\), and a secant method on it can overshoot badly. Taking the logarithm makes the function nearly linear, and the secant converges smoothly. Theo found this the hard way: a secant on \(x_{B}\) alone jumped to reflux ratios outside the bounds before it settled. The code uses a rating model of the column with 24 stages and the feed on stage 10, solving the tridiagonal stage equations by the Thomas algorithm, and wraps a secant on \(\ln x_{B}\) around it.

# A rating model of the ester column (fixed stages), then a design spec, a sensitivity table and case studies
alpha, N, NF = 2.5, 24, 10           # stages counted from the top (condenser = 0 is total), feed on stage NF
F = 63.8
def column(R, zF=0.591, D=38.2, alpha=alpha):
    """Binary, constant molar overflow, total condenser, partial reboiler = stage N. Returns xD, xB."""
    B = F - D; L, V = R * D, (R + 1) * D; Ls = L + F
    x = [zF] * (N + 1)
    for _ in range(300):
        K = [alpha / (1 + (alpha - 1) * xi) for xi in x]
        # tridiagonal balance for the light component on stages 1..N (liquid mole fractions)
        a, b, c, d = [0.0] * (N + 1), [0.0] * (N + 1), [0.0] * (N + 1), [0.0] * (N + 1)
        for j in range(1, N + 1):
            Lj = L if j < NF else Ls
            if j == N: Lj = B
            Lup = L if j - 1 < NF else Ls          # liquid from the stage above
            a[j] = Lup if j > 1 else 0.0           # x_{j-1}
            b[j] = -(Lj + V * K[j])
            c[j] = V * K[j + 1] if j < N else 0.0  # y_{j+1}
            d[j] = -F * zF if j == NF else 0.0
        # reflux liquid entering stage 1 = L * xD = L * K1 x1
        b[1] = -(L + V * K[1]) + L * K[1]
        # Thomas algorithm
        cp, dp = [0.0] * (N + 1), [0.0] * (N + 1)
        for j in range(1, N + 1):
            m = b[j] - a[j] * cp[j - 1]
            cp[j] = c[j] / m; dp[j] = (d[j] - a[j] * dp[j - 1]) / m
        xn = [0.0] * (N + 1)
        for j in range(N, 0, -1):
            xn[j] = dp[j] - (cp[j] * xn[j + 1] if j < N else 0.0)
        xn = [min(1.0, max(0.0, v)) for v in xn]
        if max(abs(p - q) for p, q in zip(x[1:], xn[1:])) < 1e-10: x = xn; break
        x = xn
    xD = K[1] * x[1]; xB = x[N]
    return xD, xB
# Design spec: find R so that the ester product contains 0.3 mol % light material (99.7 % ester)
target = 0.003
import math
# work with the logarithm of xB: the function is far straighter, so the secant method behaves
f = lambda R: math.log(column(R)[1]) - math.log(target)
R0, R1 = 1.2, 1.6
for it in range(1, 30):
    f0, f1 = f(R0), f(R1)
    R2 = R1 - f1 * (R1 - R0) / (f1 - f0)
    print(f"secant iteration {it}: R = {R2:.4f}, xB = {column(R2)[1]:.5f}")
    if abs(R2 - R1) < 1e-5: break
    R0, R1 = R1, R2
Rspec = R2
print(f"design spec met at R = {Rspec:.3f}; reflux {Rspec * 38.2:.1f} kmol/h; xD = {column(Rspec)[0]:.4f}")
print()
print("sensitivity: reflux ratio -> ester purity (mol %)")
for R in (1.2, 1.3, 1.4, 1.5, 1.7, 2.0):
    print(f"   R = {R:.1f}   purity {100 * (1 - column(R)[1]):.3f} %")
print()
print("case studies at the spec reflux")
cases = [("design feed", 0.591, alpha), ("lean feed, 55 % light", 0.55, alpha), ("rich feed, 63 % light", 0.63, alpha), ("alpha 10 % low", 0.591, alpha * 0.9)]
for name, zf, al in cases:
    xD, xB = column(Rspec, zF=zf, alpha=al)
    print(f"   {name:24s} purity {100 * (1 - xB):7.3f} %  distillate light {xD:.4f}")

secant iteration 1: R = 1.2458, xB = 0.00195
secant iteration 2: R = 1.1933, xB = 0.00516
secant iteration 3: R = 1.2226, xB = 0.00308
secant iteration 4: R = 1.2240, xB = 0.00299
secant iteration 5: R = 1.2239, xB = 0.00300
secant iteration 6: R = 1.2239, xB = 0.00300
design spec met at R = 1.224; reflux 46.8 kmol/h; xD = 0.9851

sensitivity: reflux ratio -> ester purity (mol %)
   R = 1.2   purity 99.538 %
   R = 1.3   purity 99.922 %
   R = 1.4   purity 99.970 %
   R = 1.5   purity 99.984 %
   R = 1.7   purity 99.992 %
   R = 2.0   purity 99.996 %

case studies at the spec reflux
   design feed              purity  99.700 %  distillate light 0.9851
   lean feed, 55 % light    purity  99.988 %  distillate light 0.9185
   rich feed, 63 % light    purity  97.698 %  distillate light 1.0000
   alpha 10 % low           purity  99.599 %  distillate light 0.9844

The design specification converged in six iterations to a reflux ratio of 1.224, with a reflux flow of 46.8 kmol/h. That is lower than the 1.34 from the stage-by-stage design calculation in lesson 6, because 24 stages is more than the 21 needed at 1.34. With more stages, less reflux meets the same purity.

Sensitivity analysis

A sensitivity analysis varies one or more inputs over a range and tabulates or plots chosen outputs. It does not adjust anything to meet a target; it shows how the process responds. Most simulators let the user define the varied variable, its range and step, and the outputs to record, then run the cases automatically. It is the single most useful study for understanding a process and for explaining it to others.

The sensitivity shows that purity is very sensitive to reflux near the specification. A drop from 1.224 to 1.2, only 2 %, takes purity from 99.70 to 99.54 %, off specification. Above about 1.5, extra reflux buys almost nothing. Wren draws the operational conclusion at once: running at exactly the specification reflux leaves no margin for disturbances, and the plant will run with some margin. How much margin is an economic question, answered in lesson 12.

Case studies

A case study runs the model for a set of discrete scenarios, each with its own inputs, and compares the results. Typical cases for a design are the design case, minimum and maximum throughput, summer and winter cooling water, fresh and aged catalyst, and feeds at the edges of their expected composition range. Each case uses the same model with different inputs. Simulators provide a case manager or scenario tool for this; Theo keeps a table of cases in the model log with the inputs for each.

The code runs four cases at the specification reflux. The lean feed, with 55 % light material instead of 59.1 %, gives higher purity, because there is less light material to remove. The rich feed, with 63 % light, gives 97.70 %, far off specification. The reason is instructive. The model holds the distillate rate fixed at its design value. A rich feed has more light material than that distillate can carry, so the excess must go to the bottoms, whatever the reflux. The distillate composition reaches pure light material and can do no more. This is a specification problem, not a stage or reflux problem: the plant needs to raise the distillate rate when the feed gets richer, which is what the composition control from CHE-374 does by adjusting the distillate-to-feed ratio.

The last case reduces the relative volatility by 10 %, to represent uncertainty in the thermodynamic model. The purity falls to 99.60 %, off specification. It tells Theo and Ingrid that the column has little design margin against thermodynamic uncertainty at the specification reflux. Combined with the sensitivity result, it supports designing the reflux system for a ratio well above 1.224.

Specifications that cannot be met

A design specification can fail to converge for reasons that have nothing to do with the solver. The target may be unattainable within the bounds on the manipulated variable, as the rich-feed case shows: no reflux ratio achieves 99.7 % if the distillate rate is too small. The target may be attainable only at a value outside the bounds, which the solver will hit and stop. Or the relation may not be monotonic, with two or more inputs giving the same output. A quick sensitivity analysis over the range of the manipulated variable, before setting the specification, reveals all three. Theo makes it a habit: sensitivity first, then the specification.

Design specifications should also be on variables the plant can measure and control. A purity specification on the bottoms is measured by an online analyser. A specification on a tray composition that no instrument measures may be fine for design but tells operators nothing. Where possible, Theo sets specifications that match the plant's control structure, so that the model behaves as the plant will.