Published Sep 6, 2026 · Last updated Oct 30, 2026 · 6 min · IndieRF
Part-to-Part S21 from Touchstone Files: Score a Lot
Five simulated 2.4 GHz band-pass units average 1.47 dB insertion loss at 2.45 GHz with 0.03 dB peak-to-peak spread, and the same five spread 0.59 dB at 2.2 GHz. Overlay the .s2p files and score the band the spec means.
Budget a filter from one file and you're budgeting one part. Five units of a 2.4 GHz band-pass average 1.47 dB of insertion loss at 2.45 GHz, with 0.03 dB peak to peak across the five. At 2.2 GHz, on the skirt, the same five spread 0.59 dB. Where your band edge sits decides whether the spread matters.
The five files are simulated. The IndieRF Schematic Simulator solved a 3rd-order Butterworth band-pass (centre 2.4 GHz, 20% fractional bandwidth, inductor Q 80, capacitor Q 250) five times, drawing each L within ±3% and each C within ±2%. The header of each file lists the circuit and the drawn values. Real parts spread by their own amount; the method reads yours the same way.
Open the set from the S-parameter Viewer welcome screen ("Band-pass filter, 5 units"), set the band to the spec window, and read the part-to-part stats. Files stay in the browser. The same steps work for any .s1p, .s2p, or MDIF you drop in.
The plots here are static copies of the Analysis view. Open the files on /sparam to move the marker and change the band.
Score the band the spec means
The viewer reads S21 and classifies the file as an amplifier, filter, equalizer, or generic network. For a filter, Auto X picks a padded 3 dB band from the peak, snapped to 100 MHz. On these files that's 2.1–2.7 GHz. The spec window here is 2.40–2.50 GHz. Every KPI (IL, ripple, return loss, group-delay ripple, stopband rejection) is computed over whichever band you're in.
On Unit 1, the two windows tell different stories:
| Window | IL | Ripple | RL | GD ripple |
|---|---|---|---|---|
| Auto 2.1–2.7 GHz | ||||
| Manual 2.40–2.50 GHz |
The Auto row includes the skirts. Filter IL is minus the mean in-band , so a wider band drags the mean down the edges. Return loss does the same thing: that 1.7 dB "RL" is the match at 2.1 GHz, out where the filter is supposed to reflect.
Set Manual start and stop to the spec window before you score against a datasheet. Keep Auto when you want the measured 3 dB width. Hide a file on the left to drop it from the stats.
The KPI table follows the part
Filters get f0, 3 dB BW, fractional BW, IL, ripple, RL, GD ripple, and stopband rejection, with group delay as the companion plot. Amplifiers get gain min/max/mean, flatness, match, isolation, and K/μ, with Rollett K as the companion. Equalizers get slope.
f0 and 3 dB BW come from the unpadded peak-minus-3 dB crossings, so they stay put when you shrink the analysis window. On these units f0 runs from 2.400 to 2.416 GHz and the 3 dB width from 464 to 471 MHz. IL, ripple, and RL move with the window.
Specs in the sidebar are limits on those KPIs, plus an optional in-band min/max mask. Yield is the fraction of visible files that pass every enabled spec. Enable IL max 2 dB and RL min 15 dB over 2.40–2.50 GHz and all five pass. Score the same limits over Auto 2.1–2.7 GHz and all five fail on both, because you scored the skirts.
Overlay the lot, then open Part-to-part
Manual band 2.40–2.50 GHz, marker 2.45 GHz, quantity . The table keeps the viewer's digits.
| Unit | S21 @ 2.45 GHz | IL | Ripple | RL | GD ripple |
|---|---|---|---|---|---|
| 1 | |||||
| 2 | |||||
| 3 | |||||
| 4 | |||||
| 5 |
IL agrees to 0.04 dB across the five. Return loss spans 19 to 30 dB and GD ripple 51 to 91 ps, so match and delay vary more than loss does.
The Units tab runs once two or more files carry a unit id. Name them filter_Unit-1.s2p or put DATA FOR:Unit 1 in a comment. The example files use the filename form.
At 2.45 GHz the five-unit overlay reports:
| S21 | |
|---|---|
| Mean | |
| pk-pk | |
| min (U3) | |
| max (U1) |
Call it 1.47 dB mean and 0.014 dB σ. In a receiver budget that spread is small next to the loss itself, so the mean is the number to carry.
The envelope is min / mean / max versus frequency on a common grid. An in-band min/max mask scores yield, and Cp/Cpk appears once the mask has both sides.
The envelope widens toward the band edges. Peak to peak is 0.026 dB at 2.40 GHz and 0.085 dB at 2.50 GHz. Move the marker outside the window and it keeps growing: 0.59 dB at 2.2 GHz and 0.50 dB at 2.6 GHz, with the mean loss at 3.36 and 2.98 dB. Part tolerance moves f0 over a 16 MHz range across the five, which barely moves the flat top and slides the skirts sideways. A spec that reaches into the skirt needs the edge numbers, and the worst unit, more than the mid-band mean.
Treat five parts as a lot overlay. Five samples barely know their own σ: the 95% confidence interval is about 0.6× to 2.9× the computed value, so 0.014 dB could be 0.04 dB in the population. Fine for a budget, thin for a Cpk you'd sign.
Temperature Δ is a third Analysis tab. It wants +25C or Temp: 25C in the filename or comments. These files have none; the temperature note uses a simulated equalizer set.
What to change
Keep Unit-N in the filename and drop your own set on S-parameter Viewer.
- Confirm the kind and whether Auto X is the band you meant.
- Switch Analysis to Manual for a datasheet window. Here that's 2.40–2.50 GHz, marker 2.45 GHz.
- Read the KPI table, then open Part-to-part for mean, σ, and pk-pk.
- Move the marker to each band edge and read the spread there.
- Switch quantity to or S21 group delay if that's the spec you bought. Enable a spec or a mask to score yield.
When a typical file is enough
One eval-board .s2p is the right number for a first cascade sketch. Overlay the lot when insertion loss is your margin, when your band edge sits on the filter's skirt, or when "typical" is about to go into a Monte Carlo that's supposed to be unit-to-unit. The next lot gets judged against your max spec, whatever this one did.
Questions? Contact
Discussion
No comments yet. Start the thread with a measurement, a correction, or a worked example.